Skip to main content

mnemo_engine/store/
mod.rs

1//! The Mnemo database — ties the storage engine and the memory model together.
2//!
3//! Layout of a populated file:
4//!
5//! ```text
6//!   page 0           : Header (unencrypted)
7//!   pages 1..W       : write-ahead log region
8//!   pages W..        : encrypted record runs + catalog runs + index +
9//!                      snapshot manifest (append-only)
10//! ```
11//!
12//! Durability is provided by a **write-ahead log** ([`crate::wal`]). A `flush`
13//! is one transaction: record data pages are written copy-on-write to fresh
14//! pages, then the new catalog, ANN index, and header are logged to the WAL
15//! and committed with a single fsync. A checkpoint then folds the WAL into the
16//! home pages. A crash before the commit leaves the previous state intact; a
17//! crash after it is repaired by replaying the WAL on open.
18//!
19//! Because pages are only ever appended, every past transaction's runs survive
20//! on disk. Each flush also appends an entry to a **snapshot manifest**, so
21//! [`Mnemo::restore_to`] can reinstate any past state. Stale pages — and the
22//! history along with them — are reclaimed by [`Mnemo::compact_file`].
23
24use std::fs::{File, OpenOptions};
25use std::io::{Read, Seek, SeekFrom, Write};
26use std::path::{Path, PathBuf};
27
28use fs4::fs_std::FileExt;
29use serde::{Deserialize, Serialize};
30use serde_json::{Map, Value};
31use ulid::Ulid;
32
33use crate::crypto::{self, KdfParams, SecretKind};
34use crate::embed::{self, Embedder};
35use crate::error::{MnemoError, Result};
36use crate::format::{
37    Header, RunPtr, FLAG_ENCRYPTED, PAGE_SIZE, PAYLOAD, VERSION, VERSION_V9, WRAPPED_DEK_LEN,
38};
39use crate::freelist::FreeList;
40use crate::index::{IndexConfig, IndexInfo, IndexOp, IvfPqIndex};
41#[cfg(feature = "dev-hooks")]
42use crate::keyslot::KeySlot as KeySlotEntry;
43use crate::keyslot::KeySlotRun;
44use crate::memory::{
45    self, status as essence_status, CommitBatch, CommitReport, EssenceOp, EssenceOpKind,
46    EssenceScope, EssenceSection, ExportedTurn, Filter, FilterTerm, IngestReport, KeyScope, Memory,
47    MemoryPatch, MemoryType, Metric, Scope, ScoreWeights, ThreadTurn, TurnPolicy, WriteOutcome,
48    CONTINUATION_TIMELINE_ENTRIES, ESSENCE_DEFAULT_RENDER_BYTES, ESSENCE_TRANSITIONS,
49    MNEMO_META_CONTINUES, MNEMO_META_INVALID_AT, MNEMO_META_KEY, MNEMO_META_KIND,
50    MNEMO_META_KIND_ESSENCE, MNEMO_META_KIND_THREAD_LINK, MNEMO_META_KIND_TURN, MNEMO_META_LENGTH,
51    MNEMO_META_MARKERS_ONLY, MNEMO_META_PART, MNEMO_META_REFS, MNEMO_META_ROLE, MNEMO_META_SECTION,
52    MNEMO_META_SEQ, MNEMO_META_SHA256, MNEMO_META_STATUS, MNEMO_META_STATUS_HISTORY,
53    MNEMO_META_SUPERSEDED_BY, MNEMO_META_SUPERSEDES, MNEMO_META_THREAD, MNEMO_META_TURN_ID,
54    MNEMO_META_UPDATED_AT, THREAD_TOOL_RESULT_TRUNCATE_BYTES, THREAD_TURN_CHUNK_BYTES,
55};
56pub(crate) use crate::memory::{
57    is_caller_key, is_coordination, MNEMO_META_ACKED_AT, MNEMO_META_AUTHOR,
58    MNEMO_META_CONVERSATION, MNEMO_META_DEFAULT_AGENT, MNEMO_META_DELIVERY, MNEMO_META_FROM,
59    MNEMO_META_HOPS, MNEMO_META_TO,
60};
61use crate::pager::{LockedFile, Pager};
62use crate::result_cache::{
63    self, BatchState, CacheBudget, CacheDirectoryEntry, CacheEntry, CacheFlushPolicy, CachePutOpts,
64    CacheStats, CachedValue, CounterState, DirectoryIndex, SemanticCachePutOpts,
65};
66use crate::text;
67use crate::veccol;
68use crate::wal;
69pub use coordination::{
70    slug_of, Alive, Claim, ClaimOutcome, ClaimStatus, KindCounts, Message, NewWork, Presence,
71    Stale, Task, TaskClaim, TaskLost, TaskOutcome, TaskSpec, TaskTaken, Wake,
72};
73
74mod ann;
75mod cache_api;
76mod continuation;
77mod crud;
78mod essence;
79pub(crate) use essence::format_unix_date;
80mod coordination;
81mod flush;
82#[cfg(test)]
83mod header_graft_tests;
84// HOOKS-1: the hidden test hooks exist only in test, bench and example
85// builds (the `dev-hooks` feature), never in a release build.
86#[cfg(feature = "dev-hooks")]
87mod hooks;
88mod keyed;
89mod load;
90mod maintenance;
91mod open;
92mod recall;
93mod records;
94mod shared;
95mod snapshots;
96mod threads;
97use crud::*;
98use essence::*;
99use load::*;
100pub(crate) use load::{is_file_at, is_lock_conflict};
101pub(crate) use maintenance::{create_temp, remove_dead_temps};
102pub(crate) use open::wal_header_salts;
103pub(crate) use records::plain_record;
104use records::*;
105pub use shared::SharedMnemo;
106
107/// Default initial size of the write-ahead log region, in pages (64 KiB).
108/// The region grows automatically when a transaction's control plane outgrows
109/// it, so this is a *floor* on the fresh-file footprint, not a cap on
110/// transaction size. Was 64 pages (512 KiB) in v0.1.0 — lowered because that
111/// reservation dominated small-file size (~62% of a 31-memory dogfood file).
112const DEFAULT_WAL_PAGES: u64 = 8;
113
114/// Hard floor on the WAL reservation. A single header-page flush already
115/// needs one full DATA frame plus a COMMIT frame; two pages are the bare
116/// minimum that lets the first transaction commit without immediate growth.
117const MIN_WAL_PAGES: u64 = 2;
118
119/// Default cap on retained snapshot manifest entries. Each `flush` appends
120/// one entry; without a cap, long-running processes accumulate entries
121/// forever, growing per-flush manifest-serialize cost as O(total flushes)
122/// and the manifest run size on disk in lockstep. 256 keeps roughly a
123/// week of hourly snapshots while staying tiny (256 × ~82 bytes ≈ 21 KiB
124/// of manifest). Set `MnemoConfig::max_snapshots = 0` to disable the cap.
125pub const DEFAULT_MAX_SNAPSHOTS: usize = 256;
126
127/// PERF-1: the ANN index is derived data, so `flush` persists it only every
128/// this many commits that changed it (or see
129/// [`INDEX_PERSIST_EVERY_CHANGES`], [`Mnemo::flush_index`], [`Mnemo::close`]).
130/// Between persists, only the catalog is rewritten.
131pub const INDEX_PERSIST_EVERY_COMMITS: u32 = 64;
132
133/// PERF-1: …or once this many vectors were added, changed or removed since
134/// the last persist. Records written after the persisted index are
135/// re-assigned when the index is next loaded, so this bounds that work.
136pub const INDEX_PERSIST_EVERY_CHANGES: usize = 1_000;
137
138/// FMT-9d: default weight ε of the normalised BM25 score in hybrid recall
139/// (`relevance = similarity + ε · bm25 / max bm25`).
140pub const DEFAULT_KEYWORD_WEIGHT: f32 = 0.5;
141
142/// 12.3 step 4 (v0.15): default weight of the BM25 score when it comes from
143/// a recall's own text (hybrid by default, no keywords given). Lower than
144/// [`DEFAULT_KEYWORD_WEIGHT`] because a question's common words ("how",
145/// "the") match unrelated memories: on this project's brain at 0.5 a memory
146/// with similarity 0.18 came second on "how releases are approved"; at 0.25
147/// the right fact led all three DOGFOOD-0 questions with no such intruder
148/// (`docs/design/12.3-retrieval.md`, section 5).
149pub const DEFAULT_TEXT_KEYWORD_WEIGHT: f32 = 0.25;
150
151/// FMT-9b: default [`Mnemo::set_vector_cache_budget`]: the fp16 vector
152/// column stays resident while it fits (100K × 384 dims is about 77 MB) and
153/// is read through the page cache above it. PLANNER-A, 2026-11-07.
154pub const DEFAULT_VECTOR_CACHE_BUDGET: usize = 96 * 1024 * 1024;
155
156/// FMT-9b: below this many facts a v9 file has no ANN index: an exact
157/// fp16 scan of 20K × 384 dims is about a millisecond, and the IVF+PQ
158/// codebook alone would cost more disk than it saves time.
159pub const V9_ANN_MIN_FACTS: usize = 20_000;
160
161/// FMT-9f (PLANNER-A, 2026-11-08): on v9 an unpinned recall reranks at
162/// least this many candidates per requested hit (500 for `top_k` 10), or
163/// the index's own auto depth when that's deeper (above 100K). The v9
164/// index is trained on fp16-rounded vectors, and at 20K–50K the auto
165/// depth (×10–×25) lost up to 0.03 recall@10 against v8; ×50 is within
166/// 0.005 at every measured tier. v8 is unchanged.
167pub const V9_RERANK_PER_HIT: usize = 50;
168
169/// FMT-9c: page runs longer than this are written copy-on-write straight
170/// to their home pages (and fsynced) before the commit, instead of through
171/// the WAL, on v9. The WAL then carries only the header and small runs, so
172/// it stays at its 64 KiB default instead of growing to the largest
173/// catalog or index (26 MB at 100K on v8). Safe for the same reason data
174/// pages are: the pages come from the tail or from quarantined free runs,
175/// so nothing committed references them until the header flips.
176const V9_WAL_RUN_MAX_PAGES: u64 = 2;
177
178/// FMT-9c: snapshot retention on v9: the newest 16 snapshots, plus any
179/// from the last 24 hours, with retained snapshots never pinning more than
180/// a quarter of the file (past that, the oldest go first, whatever their
181/// age). `set_max_snapshots` changes the count.
182pub const V9_DEFAULT_MAX_SNAPSHOTS: usize = 16;
183const V9_SNAPSHOT_KEEP_SECS: i64 = 24 * 3600;
184/// FMT-9c: the pin cap never drops below 128 pages (1 MiB).
185const V9_PIN_FLOOR_PAGES: u64 = 128;
186
187/// FMT-9d: the layout byte at the start of a v9 ANN index run (1: the
188/// whole IVF+PQ index, followed by delta runs). A reader treats an unknown
189/// layout as no index, so a later layout (partition-major, PERF-2) ships
190/// without a format bump.
191const V9_INDEX_LAYOUT: u8 = 1;
192
193/// FMT-9b: how many sparse vector pages one commit folds into its new
194/// pages (the vacuum). Bounded so a commit's cost stays bounded.
195const VACUUM_PAGES_PER_COMMIT: usize = 4;
196
197/// FMT-9c.2: and at most this many rows moved by the vacuum per commit (a
198/// moved row is a changed catalog row, so it lands in the commit's delta).
199/// Pages with more live rows than this stay as they are until deletes thin
200/// them out.
201const VACUUM_MAX_ROWS: usize = 64;
202
203/// FMT-9b: header flag: the file also keeps each record's float vector
204/// (`MnemoConfig::store_float_vectors`).
205const FLAG_FLOAT_VECTORS: u32 = 2;
206
207/// PERF-1: when the in-memory ANN index was last persisted, and whether the
208/// next flush must persist it.
209#[derive(Debug, Default, Clone, Copy)]
210struct IndexPersist {
211    /// Persist on the next flush (build, rebuild, drop, `flush_index`,
212    /// `close`, snapshot restore).
213    forced: bool,
214    /// Commits since the last persist that left the index dirty.
215    commits: u32,
216    /// Vectors added, re-assigned or removed since the last persist.
217    changes: usize,
218}
219
220/// Configuration for creating a new database.
221#[derive(Clone, Copy, Debug)]
222pub struct MnemoConfig {
223    /// Embedding dimensionality. Every stored vector must match this.
224    pub dimensions: usize,
225    /// Key-derivation parameters.
226    pub kdf: KdfParams,
227    /// Initial size of the WAL region in 8 KiB pages. Default is
228    /// [`DEFAULT_WAL_PAGES`] (8 pages = 64 KiB), which keeps fresh files
229    /// small; the region auto-grows when a transaction's control plane needs
230    /// more space, so raising this is only a hint for the expected steady-state
231    /// transaction size. Write-heavy databases that routinely flush large
232    /// catalogs or ANN indexes can set this higher (e.g. 64 = 512 KiB) to
233    /// avoid early grow events; tiny embedded uses can leave it at the
234    /// default. Clamped to a minimum of [`MIN_WAL_PAGES`] (2 pages).
235    pub wal_pages_initial: u64,
236    /// Maximum number of snapshot manifest entries to retain. Each `flush`
237    /// appends one entry; once the cap is hit, the *oldest* entry is dropped
238    /// from the in-memory manifest before the new one is added, so the
239    /// retained set is always the most-recent N. Defaults to
240    /// [`DEFAULT_MAX_SNAPSHOTS`] (256). Set to `0` to disable the cap and
241    /// retain every snapshot forever (the pre-v0.3 behavior).
242    ///
243    /// Pages referenced *only* by pruned snapshots stay on disk until the
244    /// next [`Mnemo::compact_file`], which reclaims them. Point-in-time
245    /// recovery via [`Mnemo::restore_to`] into a pruned `txn_id` returns
246    /// [`MnemoError::NotFound`].
247    ///
248    /// Apply to an already-open database with [`Mnemo::set_max_snapshots`].
249    pub max_snapshots: usize,
250    /// The on-disk format a new file is created in. `0` (the default)
251    /// means v9, the current format. `8` creates a v8 file, for tools that
252    /// still need to read it with v0.10 or earlier; it's kept for one
253    /// release (v0.12 makes v8 read-only). **Not a stable API.**
254    #[doc(hidden)]
255    pub format_version: u16,
256    /// FMT-9b (v9 files only; ignored for v8): also keep each memory's
257    /// full-precision vector in its record, so `get()` and `recall()`
258    /// return it exactly. Off by default: a v9 file stores vectors as
259    /// fp16 (about 1e-3 relative error) in the vector column, which recall
260    /// scores from either way. Costs `4 × dimensions` bytes per memory.
261    pub store_float_vectors: bool,
262}
263
264impl Default for MnemoConfig {
265    fn default() -> Self {
266        Self {
267            dimensions: 768,
268            kdf: KdfParams::secure(),
269            wal_pages_initial: DEFAULT_WAL_PAGES,
270            max_snapshots: DEFAULT_MAX_SNAPSHOTS,
271            format_version: 0,
272            store_float_vectors: false,
273        }
274    }
275}
276
277/// FMT-9: resolve [`MnemoConfig::format_version`] for a new file.
278fn resolve_format_version(requested: u16) -> Result<u16> {
279    match requested {
280        // FMT-9f: new files are v9; the `MNEMO_DEV_FORMAT` switch is gone.
281        0 => Ok(crate::format::DEFAULT_NEW_VERSION),
282        VERSION_V9 => Ok(requested),
283        // The v8 sunset (v0.13): v8 files open read-only and compact to v9;
284        // no new one is created.
285        v => Err(MnemoError::Invalid(format!(
286            "format_version {v} is not supported for new files: new files are v9 (pass 0 or 9)"
287        ))),
288    }
289}
290
291/// Catalog entry: maps a memory ID to its page run on disk.
292///
293/// **Schema (v5+):** carries `accessed_at` and `access_count` so `recall`
294/// can update access stats by mutating the catalog entry rather than
295/// rewriting the entire record (Phase 2.1 of the improvement plan). The
296/// values in a memory's serialized record body become a stale snapshot
297/// after the first recall; [`Mnemo::read_memory`] re-populates them from
298/// the catalog so consumers always see the live values.
299#[derive(Serialize, Deserialize, Clone, Debug)]
300struct CatalogEntry {
301    /// ULID as a raw `u128`.
302    id: u128,
303    start_page: u64,
304    page_count: u32,
305    /// Exact serialized byte length of the record.
306    len: u32,
307    deleted: bool,
308    /// Unix seconds of the most recent recall hit (or write, for fresh entries).
309    accessed_at: i64,
310    /// How many times `recall` has surfaced this memory.
311    access_count: u32,
312    /// FMT-9: the generation the record's pages were sealed under (v9 page
313    /// AAD). `0`, and left out of the encoding, on v8 files, so a v8
314    /// catalog is byte-for-byte what v0.10 writes.
315    #[serde(default, skip_serializing_if = "crate::format::is_zero_u64")]
316    gen: u64,
317    /// FMT-9b: the v9 row's tags, scoring inputs and vector location
318    /// (`None`, and left out of the encoding, on v8 files).
319    #[serde(default, skip_serializing_if = "Option::is_none")]
320    v9: Option<RowV9>,
321}
322
323/// FMT-9b: what a v9 catalog row carries beyond the v8 fields, so recall
324/// can filter and score a candidate without decrypting its record, and a
325/// cold handle can select records by kind or thread.
326#[derive(Serialize, Deserialize, Clone, Copy, Debug, Default)]
327struct RowV9 {
328    /// The vector column entry: page (`0` while it's waiting for the next
329    /// flush), slot, and the page's generation.
330    vec_page: u64,
331    vec_slot: u16,
332    vec_gen: u64,
333    /// FMT-9b.2: the record's slot on its packed page (`start_page`), or
334    /// [`REC_SPAN`] when the record owns a run of whole pages.
335    rec_slot: u16,
336    /// `ROW_KIND_*`.
337    kind: u8,
338    /// `ROW_SUPERSEDED | ROW_HAS_TTL | ROW_SHARED`.
339    flags: u8,
340    /// [`memory_type_code`].
341    memory_type: u8,
342    /// FNV-1a of `_mnemo.thread`, of the 12.1 key slot, and of `agent_id`
343    /// (`0` = none). A collision only lets a non-match through to the
344    /// decrypted check.
345    thread_hash: u32,
346    key_hash: u32,
347    agent_hash: u32,
348    created_at: i64,
349    importance: f32,
350    /// `created_at + ttl_secs` when `ROW_HAS_TTL` is set.
351    expires_at: i64,
352    /// FMT-9d: the row's BM25 document id (`0`: not keyword-indexed) and
353    /// its token count. Live rows define which documents are live, and
354    /// N and avgdl, so segments never hold lengths or tombstones.
355    doc: u32,
356    doc_len: u16,
357}
358
359/// FMT-9e: by default a thread turn is written plain and compressed once
360/// it's a week old (by its `created_at`), so a live session's recent turns
361/// read at full speed and archived history is small.
362pub const DEFAULT_TURN_COMPRESS_AFTER_SECS: i64 = 7 * 24 * 3600;
363/// Frozen v4 catalog entry shape, used only by the v4→v5 migration path in
364/// [`Mnemo::open`]. Its layout matches the catalog encoding written by
365/// pre-v5 builds (5 positional fields). Do not change this struct.
366#[derive(Serialize, Deserialize, Clone, Debug)]
367struct CatalogEntryV4 {
368    id: u128,
369    start_page: u64,
370    page_count: u32,
371    len: u32,
372    deleted: bool,
373}
374
375/// Returned by [`Mnemo::prepare_for_flush`] and consumed by [`Mnemo::flush`].
376/// Carries the already-serialized control-plane bytes through the prelude
377/// so we don't re-serialize after the lease has been persisted.
378struct FlushPrelude {
379    cat_bytes: Option<Vec<u8>>,
380    /// FMT-9c.2: `cat_bytes` is a delta run, not a base catalog.
381    cat_delta: bool,
382    /// FMT-9c.3: `idx_bytes` is an index delta run, not a whole index.
383    idx_delta: bool,
384    /// FMT-9d: this commit's BM25 segment, and the segments it replaces.
385    bm25_seg: Option<(Vec<u8>, Option<usize>)>,
386    idx_bytes: Option<Vec<u8>>,
387    /// Serialized result-cache directory bytes when `dirty_cache` is
388    /// set (Phase 10.1). `None` if the cache didn't change since the
389    /// last flush — in that case the header keeps its current
390    /// `cache_start` / `cache_pages` / `cache_len` and no fresh cache
391    /// run is written.
392    cache_bytes: Option<Vec<u8>>,
393}
394
395/// One entry in the append-only snapshot manifest. Because record, catalog,
396/// and index pages are only ever *appended*, the runs a past flush wrote are
397/// still on disk; a `Snapshot` is the set of pointers needed to reconstruct
398/// the database exactly as that flush left it.
399#[derive(Serialize, Deserialize, Clone, Debug)]
400struct Snapshot {
401    txn_id: u64,
402    created_at: i64,
403    catalog_start: u64,
404    catalog_pages: u64,
405    catalog_len: u64,
406    index_start: u64,
407    index_pages: u64,
408    index_len: u64,
409    memory_count: u64,
410    /// FMT-9: generations of the catalog and index runs (v9 only; `None`,
411    /// and left out of the encoding, on v8 files).
412    #[serde(default, skip_serializing_if = "Option::is_none")]
413    gens: Option<SnapshotGens>,
414}
415
416/// FMT-9: the generations a v9 snapshot's runs were sealed under.
417#[derive(Serialize, Deserialize, Clone, Copy, Debug, Default)]
418struct SnapshotGens {
419    catalog: u64,
420    index: u64,
421    /// FMT-9c.2: the catalog's delta-chain head at this snapshot.
422    #[serde(default)]
423    delta: (u64, u64, u64, u64),
424    /// FMT-9c.3: the index's delta-chain head at this snapshot.
425    #[serde(default)]
426    index_delta: (u64, u64, u64, u64),
427    /// FMT-9d: the BM25 manifest at this snapshot.
428    #[serde(default)]
429    bm25: (u64, u64, u64, u64),
430}
431
432impl Snapshot {
433    fn catalog_gen(&self) -> u64 {
434        self.gens.map_or(0, |g| g.catalog)
435    }
436    fn index_gen(&self) -> u64 {
437        self.gens.map_or(0, |g| g.index)
438    }
439    fn delta_head(&self) -> RunPtr {
440        self.gens.map_or(RunPtr::default(), |g| tuple_ptr(g.delta))
441    }
442    fn index_delta_head(&self) -> RunPtr {
443        self.gens
444            .map_or(RunPtr::default(), |g| tuple_ptr(g.index_delta))
445    }
446}
447
448/// A restorable point in a database's history — see [`Mnemo::snapshots`].
449#[derive(Clone, Copy, Debug)]
450pub struct SnapshotInfo {
451    /// Id of the transaction that produced this snapshot (monotonic from 1).
452    pub txn_id: u64,
453    /// When the snapshot was committed (unix seconds).
454    pub created_at: i64,
455    /// Live memory count captured in the snapshot.
456    pub memory_count: u64,
457}
458
459/// What [`Mnemo::stats`] and [`Mnemo::stats_summary`] report about restore
460/// points (INSP-1, RB-001.25).
461pub(super) struct RestorePoints {
462    pub(super) retained: usize,
463    pub(super) oldest_at: Option<i64>,
464    pub(super) pinned_bytes: u64,
465}
466
467/// A retrieval request for [`Mnemo::recall`].
468#[derive(Clone, Debug)]
469pub struct RecallRequest {
470    /// Query embedding.
471    pub query: Vec<f32>,
472    /// Maximum results to return.
473    pub top_k: usize,
474    /// Restrict to these memory types (`None` = all types).
475    pub memory_types: Option<Vec<MemoryType>>,
476    /// Restrict to a single agent's view: its own memories plus shared ones.
477    pub agent_id: Option<String>,
478    /// Similarity metric.
479    pub metric: Metric,
480    /// Multi-signal score weights.
481    pub weights: ScoreWeights,
482    /// Index override: partitions to probe (`None` = index default). Ignored
483    /// when no ANN index is present.
484    pub n_probe: Option<usize>,
485    /// Index override: candidates to rerank (`None` = index default). Ignored
486    /// when no ANN index is present.
487    pub n_rerank: Option<usize>,
488    /// Whether to update each result's `accessed_at` and `access_count` in
489    /// the catalog. Defaults to `true`, matching pre-v5 behavior.
490    ///
491    /// Set to `false` for **fully read-only recall** — useful for batch
492    /// scoring, dry-run scoring, or recalls run by introspection tooling
493    /// (e.g. `mnemo about` consumers) where you don't want the score's own
494    /// observation to perturb the database. With `false`, recall does not
495    /// dirty the catalog and the next `flush` is a no-op.
496    pub track_access: bool,
497    /// Phase 12.6 pre-scoring filter: `session_id` + AND-ed metadata
498    /// conditions applied before the candidate is scored. Default is
499    /// [`Filter::default`] (no constraints). When the ANN index is on
500    /// and this filter is selective enough to starve the shortlist
501    /// below `top_k`, Phase 12.1's widen-then-exact-scan logic kicks
502    /// in so the result still carries up to `top_k` correct hits.
503    pub filter: Filter,
504    /// 13a (13.5): the visibility levels recall returns: public and team by
505    /// default; private only when asked ([`RecallRequest::visibility`]).
506    pub visibility: crate::memory::VisibilitySet,
507    /// Phase 12.8 post-scoring budget: stop adding hits once the UTF-8
508    /// bytes of their `content` fields (not the serialized record
509    /// body) would exceed this many bytes. `None` (default) means no
510    /// budget and every hit through `top_k` is returned. If the top-1
511    /// hit is already over budget it is still returned — callers that
512    /// need an absolute cap should also pre-size `top_k`.
513    ///
514    /// Counts only the UTF-8 `content` bytes so the number is stable
515    /// across the Rust / Python / CLI / MCP surfaces; the JSON or
516    /// MessagePack envelope is excluded. Roughly 4 bytes per BPE
517    /// token for English text — a 4 KB budget comfortably holds ~1K
518    /// tokens.
519    ///
520    /// Observable through [`RecallResponse::truncated`]
521    /// ([`Mnemo::recall_budgeted`]): `true` iff at least one scored
522    /// hit was dropped by this budget, `false` otherwise.
523    pub max_bytes: Option<usize>,
524    /// FMT-9d (Phase 2, v9 files): **identifiers or exact terms to match**
525    /// (a version, a PR number, a function name) for hybrid recall. BM25
526    /// scores, divided by this query's best, are added to each candidate's
527    /// similarity with weight [`RecallRequest::keyword_weight`] before the
528    /// usual scoring. With an empty `query` vector
529    /// ([`RecallRequest::keywords_only`]), recall is keyword-only.
530    ///
531    /// **Since v0.15 (12.3 step 4), a recall by text is hybrid by default:**
532    /// with `keywords` `None`, the request's [`RecallRequest::text`] is
533    /// used for BM25, so give keywords only for identifiers the question
534    /// doesn't contain. Keywords given replace the text. **Empty keywords
535    /// (`.keywords("")`) opt out:** vector-only recall. `None` on a request
536    /// with no text is plain vector recall.
537    pub keywords: Option<String>,
538    /// 12.3 step 4 (v0.15): the query text a recall by text was built from
539    /// ([`RecallRequest::from_text`] sets it). With `keywords` `None` it's
540    /// also scored with BM25 on a v9 file (hybrid by default); a v8 file,
541    /// which has no keyword index, ignores it.
542    pub text: Option<String>,
543    /// FMT-9d: weight ε of the normalised BM25 score in hybrid recall
544    /// (default [`DEFAULT_KEYWORD_WEIGHT`]), for keywords given.
545    pub keyword_weight: f32,
546    /// 12.3 step 4 (v0.15): the weight when BM25 scores the request's
547    /// [`RecallRequest::text`] (no keywords given); default
548    /// [`DEFAULT_TEXT_KEYWORD_WEIGHT`].
549    pub text_keyword_weight: f32,
550    /// FMT-9d: BM25 term-frequency saturation `k1` (default 1.2).
551    pub bm25_k1: f32,
552    /// FMT-9d: BM25 length normalisation `b` (default 0.75).
553    pub bm25_b: f32,
554}
555
556impl RecallRequest {
557    /// A request with sensible defaults for a query embedding.
558    pub fn new(query: Vec<f32>) -> Self {
559        Self {
560            query,
561            top_k: 10,
562            memory_types: None,
563            agent_id: None,
564            metric: Metric::Cosine,
565            weights: ScoreWeights::default(),
566            n_probe: None,
567            n_rerank: None,
568            track_access: true,
569            filter: Filter::default(),
570            visibility: crate::memory::VisibilitySet::DEFAULT,
571            max_bytes: None,
572            keywords: None,
573            text: None,
574            keyword_weight: DEFAULT_KEYWORD_WEIGHT,
575            text_keyword_weight: DEFAULT_TEXT_KEYWORD_WEIGHT,
576            bm25_k1: 1.2,
577            bm25_b: 0.75,
578        }
579    }
580
581    /// 13a (13.5): the visibility levels to return (private records only
582    /// when the set includes them).
583    pub fn visibility(mut self, v: crate::memory::VisibilitySet) -> Self {
584        self.visibility = v;
585        self
586    }
587
588    /// FMT-9d (v9 files): a keyword-only request: BM25 over `keywords`
589    /// (identifiers or exact terms), no query vector, so it works without
590    /// an embedder.
591    pub fn keywords_only(keywords: impl Into<String>) -> Self {
592        Self::new(Vec::new()).keywords(keywords)
593    }
594
595    /// FMT-9d (v9 files): add identifiers or exact terms to match (hybrid
596    /// recall). They replace the request's text for BM25; `""` opts out of
597    /// keyword scoring. See the `keywords` field.
598    pub fn keywords(mut self, text: impl Into<String>) -> Self {
599        self.keywords = Some(text.into());
600        self
601    }
602
603    /// 12.3 step 4: the weight of the text's keyword score when no keywords
604    /// are given (see [`RecallRequest::text_keyword_weight`]).
605    pub fn text_keyword_weight(mut self, w: f32) -> Self {
606        self.text_keyword_weight = w;
607        self
608    }
609
610    /// 12.3 step 4: the weight of this request's keyword score: the text's
611    /// when it's the text that's scored, the keywords' otherwise.
612    pub(crate) fn effective_keyword_weight(&self) -> f32 {
613        if self.keywords.is_none() {
614            self.text_keyword_weight
615        } else {
616            self.keyword_weight
617        }
618    }
619
620    /// 12.3 step 4: the text BM25 scores for this request: the keywords
621    /// given (none if they're empty), or else the query text on a v9 file.
622    pub(crate) fn keywords_for(&self, v9: bool) -> Option<&str> {
623        match self.keywords.as_deref() {
624            Some(k) if k.trim().is_empty() => None,
625            Some(k) => Some(k),
626            None if v9 => self.text.as_deref().filter(|t| !t.trim().is_empty()),
627            None => None,
628        }
629    }
630
631    /// FMT-9d: the weight ε of the keyword score in hybrid recall.
632    pub fn keyword_weight(mut self, w: f32) -> Self {
633        self.keyword_weight = w;
634        self
635    }
636
637    /// Phase 12.6: attach a pre-scoring filter. See [`Filter`] and
638    /// [`FilterTerm`] for term shapes.
639    pub fn filter(mut self, filter: Filter) -> Self {
640        self.filter = filter;
641        self
642    }
643
644    /// Phase 12.8: cap the total UTF-8 bytes of returned `content`
645    /// fields. See [`RecallRequest::max_bytes`].
646    pub fn max_bytes(mut self, n: usize) -> Self {
647        self.max_bytes = Some(n);
648        self
649    }
650
651    /// Build a request by embedding `query_text` with the supplied
652    /// embedder (Phase 3). Since v0.15 it embeds with
653    /// [`Embedder::embed_query`] and keeps the text in
654    /// [`RecallRequest::text`], so the recall is hybrid by default (see
655    /// [`RecallRequest::keywords`]; `.keywords("")` opts out). Otherwise the
656    /// same defaults as `new`. Fails if the embedder returns an empty vector
657    /// list or an embedding failure.
658    pub fn from_text(embedder: &dyn Embedder, query_text: &str) -> Result<Self> {
659        let mut vectors = embedder.embed_query(&[query_text])?;
660        let v = vectors.pop().ok_or_else(|| {
661            MnemoError::Invalid("embedder returned no vector for the query text".into())
662        })?;
663        Ok(Self::new(v).text(query_text))
664    }
665
666    /// 12.3 step 4: set the query text (see [`RecallRequest::text`]), for
667    /// a caller that embedded it itself.
668    pub fn text(mut self, text: impl Into<String>) -> Self {
669        self.text = Some(text.into());
670        self
671    }
672    /// Set the result cap.
673    pub fn top_k(mut self, k: usize) -> Self {
674        self.top_k = k;
675        self
676    }
677    /// Restrict to specific memory types.
678    pub fn types(mut self, t: Vec<MemoryType>) -> Self {
679        self.memory_types = Some(t);
680        self
681    }
682    /// Restrict to one agent's view.
683    pub fn agent(mut self, agent_id: impl Into<String>) -> Self {
684        self.agent_id = Some(agent_id.into());
685        self
686    }
687    /// Set the similarity metric (default: cosine).
688    pub fn metric(mut self, metric: Metric) -> Self {
689        self.metric = metric;
690        self
691    }
692    /// Replace the multi-signal score weights.
693    pub fn weights(mut self, weights: ScoreWeights) -> Self {
694        self.weights = weights;
695        self
696    }
697    /// Override the number of IVF partitions probed (accuracy/speed dial).
698    pub fn n_probe(mut self, n: usize) -> Self {
699        self.n_probe = Some(n);
700        self
701    }
702    /// Override the number of candidates reranked exactly (accuracy dial).
703    pub fn n_rerank(mut self, n: usize) -> Self {
704        self.n_rerank = Some(n);
705        self
706    }
707    /// Toggle whether recall updates `accessed_at` / `access_count` on
708    /// the returned memories' catalog entries (default `true`). Pass
709    /// `false` for a fully read-only recall — see the field doc.
710    pub fn track_access(mut self, track: bool) -> Self {
711        self.track_access = track;
712        self
713    }
714}
715
716/// One scored result from [`Mnemo::recall`].
717#[derive(Clone, Debug)]
718pub struct RecallResult {
719    /// The retrieved memory.
720    pub memory: Memory,
721    /// Combined multi-signal score.
722    pub score: f32,
723    /// The bare similarity component (before other signals).
724    pub similarity: f32,
725}
726
727/// Phase 12.8 budgeted-recall envelope returned by
728/// [`Mnemo::recall_budgeted`]. Wraps the `Vec<RecallResult>` the
729/// legacy [`Mnemo::recall`] returns so callers can tell whether
730/// [`RecallRequest::max_bytes`] cut anything.
731#[derive(Clone, Debug)]
732pub struct RecallResponse {
733    /// Scored hits, best first, truncated to at most `top_k` and then
734    /// to fit the `max_bytes` budget if set.
735    pub hits: Vec<RecallResult>,
736    /// `true` iff the `max_bytes` budget dropped at least one hit
737    /// that would otherwise have been returned (i.e. the result is
738    /// shorter than what `top_k` alone would produce).
739    pub truncated: bool,
740    /// Phase 12.1 + 12.6 observability: `true` iff recall fell
741    /// through to an exact scan (ANN shortlist too narrow after
742    /// filtering, and the widen pass also came up short, so the
743    /// index was skipped for one full-corpus pass). Monitoring on
744    /// this flag is how callers detect selective-filter regressions
745    /// or supersession chains outgrowing the ANN shortlist.
746    pub used_exact_scan: bool,
747}
748
749/// Where a file's bytes and an open handle's heap go (P-0 device budget,
750/// `examples/device_budget.rs`). **Unstable diagnostic**: fields may
751/// change in any release. File figures come from the header and the
752/// catalog. RAM figures are buffer capacities, with hash maps estimated
753/// from their capacity; they cover the engine's main structures, not
754/// every allocation, and not allocator overhead.
755#[doc(hidden)]
756#[derive(Clone, Debug, Default, serde::Serialize)]
757#[cfg(feature = "dev-hooks")]
758pub struct BudgetReport {
759    pub file_bytes: u64,
760    pub page_size: u64,
761    /// Page 0.
762    pub header_bytes: u64,
763    /// Whole pages holding live records (superseded versions included).
764    pub record_page_bytes: u64,
765    /// Serialized bytes of those records; the rest of their pages is
766    /// slack.
767    pub record_payload_bytes: u64,
768    /// Pages still holding soft-deleted records (reclaimed by compact).
769    pub deleted_record_page_bytes: u64,
770    pub catalog_run_bytes: u64,
771    pub catalog_payload_bytes: u64,
772    pub index_run_bytes: u64,
773    pub index_payload_bytes: u64,
774    /// First page of the persisted index run (0 = none).
775    pub index_start_page: u64,
776    pub wal_bytes: u64,
777    pub snapshot_manifest_bytes: u64,
778    pub cache_dir_run_bytes: u64,
779    /// Everything else: superseded catalog, index and directory runs from
780    /// earlier flushes, and freed pages, until the next compact.
781    pub unreferenced_bytes: u64,
782    pub live_records: u64,
783    pub ram_catalog: u64,
784    pub ram_id_map: u64,
785    pub ram_ann_centroids: u64,
786    pub ram_ann_codebook: u64,
787    pub ram_ann_posting_ids: u64,
788    pub ram_ann_posting_codes: u64,
789    pub ram_ann_assignment: u64,
790    pub ram_cache_dir: u64,
791    pub ram_key_caches: u64,
792    /// FMT-9b: the resident fp16 vector column as allocated (rows not yet
793    /// touched by recall cost no physical memory).
794    pub ram_vector_column: u64,
795    /// Sum of the `ram_*` fields.
796    pub ram_total: u64,
797}
798
799/// Summary statistics for a database.
800#[derive(Clone, Debug)]
801pub struct Stats {
802    /// Live (non-deleted) memory count.
803    pub memories: usize,
804    /// Tombstoned entries awaiting compaction.
805    pub deleted: usize,
806    /// Embedding dimensionality.
807    pub dimensions: usize,
808    /// File size in bytes.
809    pub file_bytes: u64,
810    /// Distinct agent IDs present.
811    pub agents: Vec<String>,
812    /// Whether pages are encrypted (always true in v1).
813    pub encrypted: bool,
814    /// Creation time (unix seconds).
815    pub created_at: i64,
816    /// ANN index shape, if an index has been built.
817    pub index: Option<IndexInfo>,
818    /// Current size of the write-ahead log region, in 8 KiB pages.
819    pub wal_pages: u64,
820    /// Fsyncs issued by the most recent successful [`Mnemo::flush`] on
821    /// this handle. `0` on a fresh handle before any flush, or when
822    /// `flush()` short-circuited (nothing dirty). Phase 6 instrumentation
823    /// for the Phase 11.1 write-latency work — the planner wants this
824    /// before deciding whether cutting fsync count is a viable
825    /// durability-preserving fix.
826    pub fsyncs_last_flush: u32,
827    /// FMT-9f: the file's on-disk format (`8` or `9`). Since v0.13 a v8 file
828    /// opens read-only; `mnemo compact` converts it to v9.
829    pub format_version: u16,
830    /// MOBILE-2d: how recall scores vectors on this handle right now.
831    pub recall_mode: RecallMode,
832    /// INSP-1 (RB-001.25): the snapshots retained right now, each a
833    /// restore point for [`Mnemo::restore_to`]. Under churn the pin cap can
834    /// keep fewer than the policy's 16-and-24-hours, so this, not the
835    /// policy, is what can be restored.
836    pub snapshots_retained: usize,
837    /// When the oldest retained snapshot was committed (unix seconds):
838    /// the furthest back a restore reaches. `None` with no snapshot.
839    pub oldest_restorable_at: Option<i64>,
840    /// The bytes the retained snapshots keep from reuse (pages freed after
841    /// the oldest one). `0` on v8, which never reuses pages.
842    pub snapshot_pinned_bytes: u64,
843    /// TOK-1 (v0.15): the keyword index's tokenizer: `V2` for a file v0.15
844    /// created or rebuilt, `V1` for an older file until a compaction or
845    /// `reembed` rebuilds it, `Mixed` after a build before v0.15 wrote to
846    /// a tokenizer 2 index. `None` on v8, which has no keyword index.
847    pub keyword_tokenizer: Option<crate::KeywordTokenizer>,
848}
849
850/// [`Mnemo::stats_summary`]: the cheap statistics, read from the catalog and
851/// the header (v0.12.2, RB-002.3).
852#[derive(Clone, Debug, PartialEq, Eq)]
853pub struct StatsSummary {
854    /// Live (non-deleted) memory count.
855    pub memories: usize,
856    /// Tombstoned entries awaiting compaction.
857    pub deleted: usize,
858    /// Embedding dimensionality.
859    pub dimensions: usize,
860    /// File size in bytes.
861    pub file_bytes: u64,
862    /// The on-disk format, `8` or `9`.
863    pub format_version: u16,
864    /// Current size of the write-ahead log region, in 8 KiB pages.
865    pub wal_pages: u64,
866    /// Whether an IVF+PQ index exists (loaded, or persisted in the file).
867    pub has_index: bool,
868    /// How recall scores vectors right now.
869    pub recall_mode: RecallMode,
870    /// The bytes the v9 vector column needs to stay in memory: the live
871    /// facts × (`2 × dimensions + 4`). Recall is `ExactResident` while the
872    /// vector cache budget is at least this; `0` on v8, which has no column.
873    pub vector_column_bytes: u64,
874    /// INSP-1 (RB-001.25): the snapshots retained right now, each a
875    /// restore point for [`Mnemo::restore_to`]. Under churn the pin cap can
876    /// keep fewer than the policy's 16-and-24-hours, so this, not the
877    /// policy, is what can be restored.
878    pub snapshots_retained: usize,
879    /// When the oldest retained snapshot was committed (unix seconds):
880    /// the furthest back a restore reaches. `None` with no snapshot.
881    pub oldest_restorable_at: Option<i64>,
882    /// The bytes the retained snapshots keep from reuse (pages freed after
883    /// the oldest one). `0` on v8, which never reuses pages.
884    pub snapshot_pinned_bytes: u64,
885}
886
887/// MOBILE-2d: how recall scores vectors, from [`Mnemo::stats`].
888#[derive(Clone, Copy, Debug, PartialEq, Eq)]
889pub enum RecallMode {
890    /// An exact scan over the fp16 vector column held in memory: a v9 file
891    /// whose live facts fit [`Mnemo::set_vector_cache_budget`]. The fast
892    /// path.
893    ExactResident,
894    /// An exact scan that reads vectors through the page cache: a v9 file
895    /// whose live facts are over the budget, or a v8 file. Slower, and
896    /// slower still as the file grows; raise the budget or build an index.
897    ExactPaged,
898    /// An IVF+PQ index shortlists the candidates
899    /// ([`Mnemo::build_index`]).
900    Indexed,
901}
902
903impl RecallMode {
904    /// `exact_resident`, `exact_paged` or `indexed`.
905    pub fn as_str(&self) -> &'static str {
906        match self {
907            RecallMode::ExactResident => "exact_resident",
908            RecallMode::ExactPaged => "exact_paged",
909            RecallMode::Indexed => "indexed",
910        }
911    }
912}
913
914/// The error from [`Mnemo::try_close`]: the final flush failed, and the
915/// handle comes back with it, still open and still holding every write that
916/// wasn't flushed.
917pub struct CloseError {
918    pub(super) handle: Box<Mnemo>,
919    pub(super) error: MnemoError,
920}
921
922impl CloseError {
923    /// Why the flush failed.
924    pub fn error(&self) -> &MnemoError {
925        &self.error
926    }
927
928    /// The handle, to retry [`Mnemo::flush`] or [`Mnemo::try_close`].
929    pub fn into_inner(self) -> Mnemo {
930        *self.handle
931    }
932
933    /// The handle and the error.
934    pub fn into_parts(self) -> (Mnemo, MnemoError) {
935        (*self.handle, self.error)
936    }
937}
938
939impl std::fmt::Debug for CloseError {
940    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
941        f.debug_struct("CloseError")
942            .field("error", &self.error)
943            .finish_non_exhaustive()
944    }
945}
946
947impl std::fmt::Display for CloseError {
948    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
949        write!(f, "close failed (the handle is kept): {}", self.error)
950    }
951}
952
953impl std::error::Error for CloseError {
954    fn source(&self) -> Option<&(dyn std::error::Error + 'static)> {
955        Some(&self.error)
956    }
957}
958
959/// Result of a [`Mnemo::compact_file`] run.
960#[derive(Clone, Copy, Debug, Default)]
961pub struct CompactReport {
962    /// Live memories before compaction.
963    pub before: usize,
964    /// Live memories after compaction (expired ones dropped).
965    pub after: usize,
966    /// Live, non-expired cache entries before compaction. FIX-1
967    /// R-001.1: compaction now preserves the result cache instead of
968    /// silently dropping it.
969    pub cache_before: usize,
970    /// Live, non-expired cache entries after compaction (tombstoned
971    /// and TTL-expired ones dropped).
972    pub cache_after: usize,
973    /// Legacy onboarding or manifest memories that held an all-zero vector
974    /// (from before the zero-vector rule) and were given the sentinel unit
975    /// vector a current manifest has.
976    pub zero_vector_manifests: usize,
977    /// Other memories with an all-zero vector, carried unchanged: their
978    /// content and recall behaviour are the same as before.
979    pub zero_vector_memories: usize,
980}
981
982/// What [`Mnemo::reembed_file`] did (12.3 step 4, v0.15).
983#[derive(Clone, Debug, PartialEq, Eq)]
984pub struct ReembedReport {
985    /// Live memories embedded again: facts, superseded versions, essence
986    /// items and thread turns (expired ones are dropped, as by compaction).
987    pub memories: usize,
988    /// Manifests kept on the sentinel vector (resized to the new width).
989    pub manifests: usize,
990    /// Thread links (`_mnemo.kind = "thread_link"`) kept on the sentinel
991    /// vector (resized to the new width), R-008.3.
992    pub thread_links: usize,
993    /// Result-cache entries copied (the engine's own aren't counted).
994    pub cache_entries: usize,
995    /// Semantic cache entries dropped because their vector has the old
996    /// width (only when the width changed).
997    pub cache_dropped: usize,
998    /// The identity the file recorded before, if any.
999    pub from: Option<crate::EmbedderIdentity>,
1000    /// The identity it records now.
1001    pub to: crate::EmbedderIdentity,
1002}
1003
1004/// What [`Mnemo::verify_report`] checked.
1005#[derive(Clone, Debug, PartialEq, Eq)]
1006pub struct VerifyReport {
1007    /// Records decrypted and decoded: live memories and cache entries (the
1008    /// engine's own entries aren't counted).
1009    pub records: usize,
1010    /// Live memories whose vector is all zeros (from before the zero-vector
1011    /// rule). They score 0.0 on every vector query.
1012    pub zero_vector_memories: usize,
1013    /// TOK-1: the keyword index's tokenizer (see
1014    /// [`Stats::keyword_tokenizer`]). `Mixed` is worth a compaction.
1015    pub keyword_tokenizer: Option<crate::KeywordTokenizer>,
1016}
1017
1018/// An encrypted, single-file agent memory database.
1019pub struct Mnemo {
1020    /// 12.4: the lock layer's note for a handle opened through it for a
1021    /// long operation ([`Mnemo::open_exclusive`], compaction): its
1022    /// `lock-info` line, removed when the handle drops. Declared first so
1023    /// it drops before `pager`, whose file's unlock comes after it.
1024    lock_note: Option<crate::shared::Held>,
1025    /// 12.3: what this handle has to tell its owner once, such as a
1026    /// passphrase on raw-key KDF parameters ([`Mnemo::notices`]).
1027    notices: Vec<String>,
1028    pager: Pager,
1029    header: Header,
1030    catalog: Vec<CatalogEntry>,
1031    index: IdIndex,
1032    /// The path this handle opened, made absolute at open (so a later
1033    /// change of the process's directory doesn't change what it names).
1034    /// `ensure_not_replaced` checks that it still names the open file.
1035    path: PathBuf,
1036    dimensions: usize,
1037    kdf: KdfParams,
1038    /// Set whenever the catalog changes; drives whether `flush` rewrites it.
1039    dirty_catalog: bool,
1040    /// Optional IVF+PQ approximate-nearest-neighbour index.
1041    ann: Option<IvfPqIndex>,
1042    /// Set whenever the in-memory ANN index differs from the persisted one.
1043    /// PERF-1: `flush` persists it only per [`IndexPersist`]'s policy.
1044    dirty_index: bool,
1045    /// PERF-1: index persistence policy state.
1046    index_persist: IndexPersist,
1047    /// PERF-1 step 3: the header points at a persisted ANN index that
1048    /// hasn't been loaded yet. It loads on first use (`ensure_ann`).
1049    ann_unloaded: bool,
1050    /// Append-only manifest of committed snapshots, oldest first. Capped
1051    /// at [`Mnemo::max_snapshots`] entries on every flush — the oldest
1052    /// entries are pruned first.
1053    manifest: Vec<Snapshot>,
1054    /// Maximum manifest entries to retain across flushes. `0` disables
1055    /// the cap (retain forever). Defaults to [`DEFAULT_MAX_SNAPSHOTS`].
1056    max_snapshots: usize,
1057    /// If `true`, this handle was opened via [`Mnemo::open_read_only`] and
1058    /// every mutating API returns [`MnemoError::ReadOnly`]. The underlying
1059    /// file was opened without write permission and holds an OS shared
1060    /// lock; the OS would reject a write() anyway, but gating in Rust
1061    /// gives callers a clean, typed error.
1062    read_only: bool,
1063    /// Result-cache directory (Phase 10.1). See [`crate::result_cache`].
1064    /// Loaded from `header.cache_start` on open; persisted on `flush`
1065    /// via the same seal_run pipeline as catalog/index/manifest.
1066    cache_dir: Vec<CacheDirectoryEntry>,
1067    /// In-memory index over `cache_dir` for O(1) `(namespace, key_hash)`
1068    /// lookups. Rebuilt on open + kept in sync on `cache_put`/`delete`.
1069    cache_idx: DirectoryIndex,
1070    /// Set whenever `cache_dir` changes; drives whether `flush` rewrites
1071    /// the cache directory run.
1072    dirty_cache: bool,
1073    /// Per-namespace eviction budgets. Missing entries fall back to
1074    /// [`CacheBudget::default`]. Not persisted (per-handle policy).
1075    cache_budgets: std::collections::HashMap<String, CacheBudget>,
1076    /// Flush policy for cache mutations. Memory writes ignore this and
1077    /// always behave as `Strict`.
1078    cache_flush_policy: CacheFlushPolicy,
1079    /// Bookkeeping for `CacheFlushPolicy::Batched` — pending mutation
1080    /// count and first-dirty timestamp.
1081    cache_batch: BatchState,
1082    /// In-memory hit/miss/eviction counters (reset on reopen).
1083    cache_counters: CounterState,
1084    /// Optional handle-scoped embedder (Phase 3, PR 10a). Registered
1085    /// via [`Mnemo::set_embedder`]; consulted by [`Mnemo::remember_text`]
1086    /// and [`RecallRequest::from_text`]. `None` means text-side APIs
1087    /// return [`MnemoError::Invalid`] with a "no embedder" message.
1088    /// Not persisted — the handle is re-registered after every open.
1089    embedder: Option<std::sync::Arc<dyn Embedder>>,
1090    /// Fsyncs issued by the most recent `flush()` call. Snapshotted
1091    /// from the thread-local fsync counter in `lib.rs::fsync_counter`:
1092    /// `flush()` reads it on entry, does its work (each `sync_all` the
1093    /// pager / wal / recovery path makes bumps the counter), and
1094    /// stores the delta here. `0` on a fresh handle before any flush,
1095    /// or on a read-only handle where `flush()` short-circuits.
1096    /// Exposed through [`Stats::fsyncs_last_flush`] for diagnostics —
1097    /// the Phase 6 planner wants this before deciding whether cutting
1098    /// fsync count is a viable write-latency win.
1099    fsyncs_last_flush: u32,
1100    /// FIX-1 R-001.8 part 2: lazy cache of the distinct agent IDs
1101    /// across every live memory, built once on first `stats()` call
1102    /// per handle. `remember()` adds to it in O(1); `delete()` can't
1103    /// cheaply tell whether the removed memory was the last of its
1104    /// agent (the agent lives in the encrypted body, not the catalog
1105    /// — pre-v9 format), so it just invalidates the set and the next
1106    /// `stats()` call rebuilds. Pre-FIX-1 every `stats()` call
1107    /// decrypted every live memory to collect agents, turning an
1108    /// otherwise-cheap diagnostic into an O(N) decrypt.
1109    agent_set: Option<std::collections::BTreeSet<String>>,
1110    /// Phase 12.1 lazy caches, built together on the first keyed op this
1111    /// handle sees (`remember_keyed`, `update` that touches `_mnemo.key`,
1112    /// or any `history(...)` call). Both are maintained incrementally
1113    /// after that. `None` → not built yet; `Some(…)` → authoritative.
1114    ///
1115    /// * `key_index`: `(scope_owner, key) → ULID bytes` of the **live**
1116    ///   record for each keyed slot. `scope_owner` is `Some(agent_id)`
1117    ///   for a `Private` scope slot and `None` for a `Shared` one, so a
1118    ///   private write never collides with a shared one.
1119    /// * `superseded`: the ULID bytes of every record that has been
1120    ///   superseded (`_mnemo.superseded_by` set on it). `recall` and
1121    ///   `about` skip these before any decrypt.
1122    ///
1123    /// Building both costs one scan over live records (~1s at 100K with
1124    /// hardware AES) and happens at most once per handle. A format bump
1125    /// to v9 moves the key into the catalog row so the scan goes away.
1126    key_index: Option<std::collections::HashMap<KeySlot, u128>>,
1127    superseded: Option<std::collections::HashSet<u128>>,
1128    /// R-004.2: thread layers + onboarding ids, built by the same lazy scan.
1129    layers: Option<LayerCache>,
1130    /// REKEY-1 (RB-004.2): what the last successful rekey couldn't scrub
1131    /// once the new secret was in force (see [`Mnemo::rekey_cleanup_error`]).
1132    rekey_cleanup: Option<String>,
1133    /// 12.4 shared mode: access statistics from reads made under the
1134    /// shared lock (recall hits, cache hits), applied with this handle's
1135    /// next commit. `None` on every other handle.
1136    access_delta: Option<shared::AccessDelta>,
1137    /// 12.4 shared mode: the transaction id of the commit `flush` had
1138    /// reached the WAL with, when it fails after that point. A shared write
1139    /// uses it to tell a lost write from one that may be in the file.
1140    commit_point: Option<u64>,
1141    /// 12.4 shared mode (RB-007.5): a `flush` since the window started
1142    /// committed. Set by every committing flush; a shared write window
1143    /// clears it first, and then knows that part of the window is in the
1144    /// file when a later step fails.
1145    window_committed: bool,
1146    /// 12.4 shared mode: this handle is a shared handle's state, so a
1147    /// `rekey` inside a window is refused (`SharedMnemo::rekey` does it and
1148    /// updates the kept secret).
1149    shared: bool,
1150    /// FMT-9e: thread turns are compressed once older than this many
1151    /// seconds (`0`: at write). Per handle, not persisted.
1152    turn_compress_after: i64,
1153    /// 13a: stamped as `_mnemo.author` on each new record this handle
1154    /// writes ([`Mnemo::set_author`]).
1155    author: Option<String>,
1156    /// 13a (13.3): this handle's process holds the author's presence lock
1157    /// (a shared handle's [`SharedMnemo::hold_presence`]), so its claims
1158    /// may lapse early when it dies.
1159    presence_held: bool,
1160    /// 13a: seconds added to the clock coordination reads (tests only, via
1161    /// a dev-hooks setter).
1162    clock_offset: i64,
1163    /// RB-013.1: the ids of coordination records whose v9 row an older
1164    /// build re-tagged as a fact (its `compact` rewrites every row), found
1165    /// once per handle by decrypting the `episodic` fact rows.
1166    legacy_coordination: Option<std::collections::HashSet<u128>>,
1167    /// FMT-9 (a): the committed free list (v9; always empty on v8). Pages
1168    /// allocated from it since the last commit are already removed.
1169    free: FreeList,
1170    /// FMT-9 (a): runs this transaction stopped referencing (overwritten
1171    /// or deleted records, tombstoned cache entries, a relocated WAL
1172    /// region). The next commit adds them to the free list, freed at its
1173    /// own transaction id.
1174    pending_free: Vec<(u64, u64)>,
1175    /// FMT-9 (a): the free list changed (an allocation took pages from it)
1176    /// and must be rewritten by the next commit.
1177    free_dirty: bool,
1178    /// FMT-9 (b): the key-slot run (v9; empty until Phase 13.2 adds slots).
1179    key_slots: KeySlotRun,
1180    /// FMT-9 (b): `key_slots` changed since the last commit.
1181    dirty_keys: bool,
1182    /// FMT-9: this writable session has burned a transaction id with a
1183    /// durable header-only commit (`ensure_gen_lease`), so the generations
1184    /// its pages carry can't repeat ones a crashed earlier session used.
1185    gen_leased: bool,
1186    /// FMT-9b: the fp16 vector column (v9; unused on v8).
1187    col: VecColumn,
1188    /// FMT-9b.2: packed record pages (v9; unused on v8).
1189    recs: RecPages,
1190    /// FMT-9d: the BM25 keyword index (v9; unused on v8).
1191    bm25: Bm25State,
1192    /// FMT-9d: test hook: write this layout byte instead of the current
1193    /// one (to exercise the unknown-layout path).
1194    index_layout_override: Option<u8>,
1195    /// FMT-9d: during one recall, each keyword match's normalised BM25
1196    /// score, by id.
1197    kw: Option<std::collections::HashMap<u128, f32>>,
1198    /// FMT-9c.2: catalog rows changed since the last commit (v9 writes
1199    /// them as a delta run instead of the whole catalog).
1200    dirty_rows: std::collections::BTreeSet<usize>,
1201    /// FMT-9c.2: the delta runs on top of the base catalog, oldest first,
1202    /// and how many rows they hold in all.
1203    delta_chain: Vec<RunPtr>,
1204    delta_rows: usize,
1205    /// FMT-9c.3: the ANN index's delta runs on top of its base run, oldest
1206    /// first, and how many changes they hold in all.
1207    index_chain: Vec<RunPtr>,
1208    index_chain_ops: usize,
1209    /// FMT-9c.3: the next commit writes the whole index (after a build,
1210    /// rebuild or restore).
1211    index_rebase: bool,
1212    /// FMT-9c.3: this flush's index changes, when it writes them as a
1213    /// delta run.
1214    index_delta_ops: Option<Vec<IndexOp>>,
1215    /// FMT-9c.2: the next commit writes a whole base catalog (after a
1216    /// restore, or when the chain has grown too long).
1217    catalog_rewrite: bool,
1218    /// FMT-9c: the snapshot count was set explicitly (config or
1219    /// `set_max_snapshots`), which replaces v9's default retention policy.
1220    snapshot_cap_explicit: bool,
1221    /// FMT-9c: this flush wrote big runs straight to their home pages,
1222    /// which must be fsynced before the commit.
1223    direct_writes: bool,
1224    /// FMT-9b: keep the column resident while it fits this many bytes.
1225    vector_cache_budget: usize,
1226}
1227
1228/// Composite key for the lazy `key_index` on `Mnemo`. A slot is identified
1229/// by `(scope_owner, key)`:
1230/// * `Scope::Private` writes carry the owning agent → `Some(agent_id)`.
1231/// * `Scope::Shared` writes are per-file → `None`.
1232///
1233/// A private write never supersedes a shared one, and the reverse is also
1234/// true. Scope lives on `Memory.scope`, agent on `Memory.agent_id`.
1235#[derive(Debug, Clone, PartialEq, Eq, Hash)]
1236pub(crate) struct KeySlot {
1237    /// `Some(agent_id)` for `Scope::Private`, `None` for `Scope::Shared`.
1238    pub(crate) scope_owner: Option<String>,
1239    /// The caller-supplied key string, exactly as given (no case-fold).
1240    pub(crate) key: String,
1241}
1242
1243/// True if a memory's metadata marks it as the canonical onboarding manifest
1244/// (`metadata.topic == "manifest"`). Used by [`Mnemo::about`] to hoist the
1245/// manifest to the top of the briefing regardless of importance ordering.
1246fn is_manifest(m: &Memory) -> bool {
1247    m.metadata
1248        .get("topic")
1249        .and_then(|v| v.as_str())
1250        .map(|s| s.eq_ignore_ascii_case("manifest"))
1251        .unwrap_or(false)
1252}
1253
1254/// PERF-1 step 4: ULID → catalog row, as an open-addressing table of `u32`
1255/// row numbers: 4 bytes a slot, about 5 bytes an entry, against about 43
1256/// for the `HashMap<u128, usize>` it replaces. A probe compares against the
1257/// catalog's own id at that row, so the ids aren't stored twice. Rows are
1258/// never removed (a delete only tombstones the catalog entry), so there's
1259/// no delete path.
1260#[derive(Debug, Clone, Default)]
1261struct IdIndex {
1262    slots: Vec<u32>,
1263    len: usize,
1264}
1265
1266const NO_ROW: u32 = u32::MAX;
1267
1268impl IdIndex {
1269    fn build(catalog: &[CatalogEntry]) -> IdIndex {
1270        let mut t = IdIndex {
1271            slots: vec![NO_ROW; crate::index::table_capacity_for(catalog.len())],
1272            len: 0,
1273        };
1274        for (row, e) in catalog.iter().enumerate() {
1275            t.insert(e.id, row, catalog);
1276        }
1277        t
1278    }
1279
1280    /// The catalog row holding `id`, if any.
1281    fn get(&self, id: u128, catalog: &[CatalogEntry]) -> Option<usize> {
1282        if self.slots.is_empty() {
1283            return None;
1284        }
1285        let cap = self.slots.len();
1286        let mut i = crate::index::slot_home(id, cap);
1287        loop {
1288            let row = self.slots[i];
1289            if row == NO_ROW {
1290                return None;
1291            }
1292            if catalog[row as usize].id == id {
1293                return Some(row as usize);
1294            }
1295            i = crate::index::next_slot(i, cap);
1296        }
1297    }
1298
1299    /// Map `id` to `row`. `catalog` must hold every row already in the
1300    /// table (the new row may not be pushed yet).
1301    fn insert(&mut self, id: u128, row: usize, catalog: &[CatalogEntry]) {
1302        if (self.len + 1) * 4 > self.slots.len() * 3 {
1303            let cap = (self.slots.len() * 2).max(crate::index::table_capacity_for(self.len + 1));
1304            let old = std::mem::replace(&mut self.slots, vec![NO_ROW; cap]);
1305            self.len = 0;
1306            for r in old.into_iter().filter(|&r| r != NO_ROW) {
1307                self.insert(catalog[r as usize].id, r as usize, catalog);
1308            }
1309        }
1310        let cap = self.slots.len();
1311        let mut i = crate::index::slot_home(id, cap);
1312        loop {
1313            let r = self.slots[i];
1314            if r == NO_ROW {
1315                self.slots[i] = row as u32;
1316                self.len += 1;
1317                return;
1318            }
1319            if catalog[r as usize].id == id {
1320                self.slots[i] = row as u32;
1321                return;
1322            }
1323            i = crate::index::next_slot(i, cap);
1324        }
1325    }
1326
1327    #[cfg(feature = "dev-hooks")]
1328    fn heap_bytes(&self) -> usize {
1329        self.slots.capacity() * std::mem::size_of::<u32>()
1330    }
1331}
1332
1333/// Build the ULID → catalog-row lookup.
1334fn build_id_index(catalog: &[CatalogEntry]) -> IdIndex {
1335    IdIndex::build(catalog)
1336}
1337
1338// ---------------------------------------------------------------------------
1339// Phase 12.7a helpers used by `ingest_thread`.
1340// ---------------------------------------------------------------------------
1341
1342/// PERF-1 / FMT-9: was this catalog row written after the persisted index
1343/// run? v8 allocates pages monotonically, so a record past the run's first
1344/// page is newer. v9 reuses pages, so it compares generations instead.
1345fn written_after_index(e: &CatalogEntry, v9: bool, index_start: u64, index_gen: u64) -> bool {
1346    if v9 {
1347        // `start_page` 0: packed at the next flush, so certainly newer.
1348        e.start_page == 0 || e.gen > index_gen
1349    } else {
1350        e.start_page > index_start
1351    }
1352}
1353
1354/// Whether a record belongs in the ANN index: facts only. Thread-archive
1355/// turns, essence items and thread links are read by kind and scope, never
1356/// by similarity, so they stay out (R-004.2).
1357/// v0.16.1: a thread turn whose content is only transcript markers
1358/// ([`MNEMO_META_MARKERS_ONLY`]), by its flag or, for one saved before the
1359/// flag, by its content.
1360pub(super) fn is_markers_only_turn(m: &Memory) -> bool {
1361    meta_str(m, MNEMO_META_KIND) == Some(MNEMO_META_KIND_TURN)
1362        && (m.metadata.get(MNEMO_META_MARKERS_ONLY) == Some(&Value::Bool(true))
1363            || crate::transcript::is_marker_only(&m.content))
1364}
1365
1366#[cfg(test)]
1367#[test]
1368fn a_marker_only_turn_saved_before_the_flag_is_recognised_by_its_content() {
1369    let turn = |content: &str| {
1370        let mut m = Memory::new(content, MemoryType::Episodic, vec![1.0]);
1371        m.metadata.insert(
1372            MNEMO_META_KIND.into(),
1373            Value::String(MNEMO_META_KIND_TURN.into()),
1374        );
1375        m
1376    };
1377    assert!(is_markers_only_turn(&turn("[thinking]")));
1378    assert!(is_markers_only_turn(&turn(
1379        "[tool_use name=mcp__mnemo__remember]"
1380    )));
1381    assert!(!is_markers_only_turn(&turn(
1382        "staging-7\n[tool_use name=Bash]"
1383    )));
1384    // Only turns: a fact that reads like a marker is a fact.
1385    assert!(!is_markers_only_turn(&Memory::new(
1386        "[thinking]",
1387        MemoryType::Semantic,
1388        vec![1.0]
1389    )));
1390}
1391
1392/// 13a: a caller's `_mnemo.*` keys: [`crate::memory::CALLER_KEYS`] are
1393/// allowed and checked; any other is refused.
1394pub(super) fn check_caller_reserved(metadata: &Map<String, Value>) -> Result<()> {
1395    for (k, v) in metadata {
1396        if !k.starts_with(crate::memory::MNEMO_META_NS) {
1397            continue;
1398        }
1399        if !is_caller_key(k) {
1400            return Err(crud::reserved_key_error(k));
1401        }
1402        crate::memory::check_caller_key(k, v).map_err(MnemoError::Invalid)?;
1403    }
1404    Ok(())
1405}
1406
1407fn is_ann_kind(m: &Memory) -> bool {
1408    !matches!(
1409        m.metadata.get(MNEMO_META_KIND).and_then(|v| v.as_str()),
1410        Some(MNEMO_META_KIND_TURN)
1411            | Some(MNEMO_META_KIND_ESSENCE)
1412            | Some(MNEMO_META_KIND_THREAD_LINK)
1413    ) && !is_coordination(m)
1414}