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