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};
73pub use findings::{valid_finding_id, Finding, FindingFilter, FindingSpec};
74
75mod ann;
76mod cache_api;
77mod continuation;
78mod crud;
79mod essence;
80pub(crate) use essence::format_unix_date;
81mod coordination;
82mod findings;
83mod flush;
84#[cfg(test)]
85mod header_graft_tests;
86// HOOKS-1: the hidden test hooks exist only in test, bench and example
87// builds (the `dev-hooks` feature), never in a release build.
88#[cfg(feature = "dev-hooks")]
89mod hooks;
90mod keyed;
91mod load;
92mod maintenance;
93mod open;
94mod recall;
95mod records;
96mod shared;
97mod snapshots;
98mod threads;
99use crud::*;
100use essence::*;
101use load::*;
102pub(crate) use load::{is_file_at, is_lock_conflict};
103pub(crate) use maintenance::{create_temp, remove_dead_temps};
104pub(crate) use open::wal_header_salts;
105pub(crate) use records::plain_record;
106use records::*;
107pub use shared::SharedMnemo;
108
109/// Default initial size of the write-ahead log region, in pages (64 KiB).
110/// The region grows automatically when a transaction's control plane outgrows
111/// it, so this is a *floor* on the fresh-file footprint, not a cap on
112/// transaction size. Was 64 pages (512 KiB) in v0.1.0 — lowered because that
113/// reservation dominated small-file size (~62% of a 31-memory dogfood file).
114const DEFAULT_WAL_PAGES: u64 = 8;
115
116/// Hard floor on the WAL reservation. A single header-page flush already
117/// needs one full DATA frame plus a COMMIT frame; two pages are the bare
118/// minimum that lets the first transaction commit without immediate growth.
119const MIN_WAL_PAGES: u64 = 2;
120
121/// Default cap on retained snapshot manifest entries. Each `flush` appends
122/// one entry; without a cap, long-running processes accumulate entries
123/// forever, growing per-flush manifest-serialize cost as O(total flushes)
124/// and the manifest run size on disk in lockstep. 256 keeps roughly a
125/// week of hourly snapshots while staying tiny (256 × ~82 bytes ≈ 21 KiB
126/// of manifest). Set `MnemoConfig::max_snapshots = 0` to disable the cap.
127pub const DEFAULT_MAX_SNAPSHOTS: usize = 256;
128
129/// PERF-1: the ANN index is derived data, so `flush` persists it only every
130/// this many commits that changed it (or see
131/// [`INDEX_PERSIST_EVERY_CHANGES`], [`Mnemo::flush_index`], [`Mnemo::close`]).
132/// Between persists, only the catalog is rewritten.
133pub const INDEX_PERSIST_EVERY_COMMITS: u32 = 64;
134
135/// PERF-1: …or once this many vectors were added, changed or removed since
136/// the last persist. Records written after the persisted index are
137/// re-assigned when the index is next loaded, so this bounds that work.
138pub const INDEX_PERSIST_EVERY_CHANGES: usize = 1_000;
139
140/// FMT-9d: default weight ε of the normalised BM25 score in hybrid recall
141/// (`relevance = similarity + ε · bm25 / max bm25`).
142pub const DEFAULT_KEYWORD_WEIGHT: f32 = 0.5;
143
144/// 12.3 step 4 (v0.15): default weight of the BM25 score when it comes from
145/// a recall's own text (hybrid by default, no keywords given). Lower than
146/// [`DEFAULT_KEYWORD_WEIGHT`] because a question's common words ("how",
147/// "the") match unrelated memories: on this project's brain at 0.5 a memory
148/// with similarity 0.18 came second on "how releases are approved"; at 0.25
149/// the right fact led all three DOGFOOD-0 questions with no such intruder
150/// (`docs/design/12.3-retrieval.md`, section 5).
151pub const DEFAULT_TEXT_KEYWORD_WEIGHT: f32 = 0.25;
152
153/// FMT-9b: default [`Mnemo::set_vector_cache_budget`]: the fp16 vector
154/// column stays resident while it fits (100K × 384 dims is about 77 MB) and
155/// is read through the page cache above it. PLANNER-A, 2026-11-07.
156pub const DEFAULT_VECTOR_CACHE_BUDGET: usize = 96 * 1024 * 1024;
157
158/// FMT-9b: below this many facts a v9 file has no ANN index: an exact
159/// fp16 scan of 20K × 384 dims is about a millisecond, and the IVF+PQ
160/// codebook alone would cost more disk than it saves time.
161pub const V9_ANN_MIN_FACTS: usize = 20_000;
162
163/// FMT-9f (PLANNER-A, 2026-11-08): on v9 an unpinned recall reranks at
164/// least this many candidates per requested hit (500 for `top_k` 10), or
165/// the index's own auto depth when that's deeper (above 100K). The v9
166/// index is trained on fp16-rounded vectors, and at 20K–50K the auto
167/// depth (×10–×25) lost up to 0.03 recall@10 against v8; ×50 is within
168/// 0.005 at every measured tier. v8 is unchanged.
169pub const V9_RERANK_PER_HIT: usize = 50;
170
171/// FMT-9c: page runs longer than this are written copy-on-write straight
172/// to their home pages (and fsynced) before the commit, instead of through
173/// the WAL, on v9. The WAL then carries only the header and small runs, so
174/// it stays at its 64 KiB default instead of growing to the largest
175/// catalog or index (26 MB at 100K on v8). Safe for the same reason data
176/// pages are: the pages come from the tail or from quarantined free runs,
177/// so nothing committed references them until the header flips.
178const V9_WAL_RUN_MAX_PAGES: u64 = 2;
179
180/// FMT-9c: snapshot retention on v9: the newest 16 snapshots, plus any
181/// from the last 24 hours, with retained snapshots never pinning more than
182/// a quarter of the file (past that, the oldest go first, whatever their
183/// age). `set_max_snapshots` changes the count.
184pub const V9_DEFAULT_MAX_SNAPSHOTS: usize = 16;
185const V9_SNAPSHOT_KEEP_SECS: i64 = 24 * 3600;
186/// FMT-9c: the pin cap never drops below 128 pages (1 MiB).
187const V9_PIN_FLOOR_PAGES: u64 = 128;
188
189/// FMT-9d: the layout byte at the start of a v9 ANN index run (1: the
190/// whole IVF+PQ index, followed by delta runs). A reader treats an unknown
191/// layout as no index, so a later layout (partition-major, PERF-2) ships
192/// without a format bump.
193const V9_INDEX_LAYOUT: u8 = 1;
194
195/// FMT-9b: how many sparse vector pages one commit folds into its new
196/// pages (the vacuum). Bounded so a commit's cost stays bounded.
197const VACUUM_PAGES_PER_COMMIT: usize = 4;
198
199/// FMT-9c.2: and at most this many rows moved by the vacuum per commit (a
200/// moved row is a changed catalog row, so it lands in the commit's delta).
201/// Pages with more live rows than this stay as they are until deletes thin
202/// them out.
203const VACUUM_MAX_ROWS: usize = 64;
204
205/// FMT-9b: header flag: the file also keeps each record's float vector
206/// (`MnemoConfig::store_float_vectors`).
207const FLAG_FLOAT_VECTORS: u32 = 2;
208
209/// PERF-1: when the in-memory ANN index was last persisted, and whether the
210/// next flush must persist it.
211#[derive(Debug, Default, Clone, Copy)]
212struct IndexPersist {
213 /// Persist on the next flush (build, rebuild, drop, `flush_index`,
214 /// `close`, snapshot restore).
215 forced: bool,
216 /// Commits since the last persist that left the index dirty.
217 commits: u32,
218 /// Vectors added, re-assigned or removed since the last persist.
219 changes: usize,
220}
221
222/// Configuration for creating a new database.
223#[derive(Clone, Copy, Debug)]
224pub struct MnemoConfig {
225 /// Embedding dimensionality. Every stored vector must match this.
226 pub dimensions: usize,
227 /// Key-derivation parameters.
228 pub kdf: KdfParams,
229 /// Initial size of the WAL region in 8 KiB pages. Default is
230 /// [`DEFAULT_WAL_PAGES`] (8 pages = 64 KiB), which keeps fresh files
231 /// small; the region auto-grows when a transaction's control plane needs
232 /// more space, so raising this is only a hint for the expected steady-state
233 /// transaction size. Write-heavy databases that routinely flush large
234 /// catalogs or ANN indexes can set this higher (e.g. 64 = 512 KiB) to
235 /// avoid early grow events; tiny embedded uses can leave it at the
236 /// default. Clamped to a minimum of [`MIN_WAL_PAGES`] (2 pages).
237 pub wal_pages_initial: u64,
238 /// Maximum number of snapshot manifest entries to retain. Each `flush`
239 /// appends one entry; once the cap is hit, the *oldest* entry is dropped
240 /// from the in-memory manifest before the new one is added, so the
241 /// retained set is always the most-recent N. Defaults to
242 /// [`DEFAULT_MAX_SNAPSHOTS`] (256). Set to `0` to disable the cap and
243 /// retain every snapshot forever (the pre-v0.3 behavior).
244 ///
245 /// Pages referenced *only* by pruned snapshots stay on disk until the
246 /// next [`Mnemo::compact_file`], which reclaims them. Point-in-time
247 /// recovery via [`Mnemo::restore_to`] into a pruned `txn_id` returns
248 /// [`MnemoError::NotFound`].
249 ///
250 /// Apply to an already-open database with [`Mnemo::set_max_snapshots`].
251 pub max_snapshots: usize,
252 /// The on-disk format a new file is created in. `0` (the default)
253 /// means v9, the current format. `8` creates a v8 file, for tools that
254 /// still need to read it with v0.10 or earlier; it's kept for one
255 /// release (v0.12 makes v8 read-only). **Not a stable API.**
256 #[doc(hidden)]
257 pub format_version: u16,
258 /// FMT-9b (v9 files only; ignored for v8): also keep each memory's
259 /// full-precision vector in its record, so `get()` and `recall()`
260 /// return it exactly. Off by default: a v9 file stores vectors as
261 /// fp16 (about 1e-3 relative error) in the vector column, which recall
262 /// scores from either way. Costs `4 × dimensions` bytes per memory.
263 pub store_float_vectors: bool,
264}
265
266impl Default for MnemoConfig {
267 fn default() -> Self {
268 Self {
269 dimensions: 768,
270 kdf: KdfParams::secure(),
271 wal_pages_initial: DEFAULT_WAL_PAGES,
272 max_snapshots: DEFAULT_MAX_SNAPSHOTS,
273 format_version: 0,
274 store_float_vectors: false,
275 }
276 }
277}
278
279/// FMT-9: resolve [`MnemoConfig::format_version`] for a new file.
280fn resolve_format_version(requested: u16) -> Result<u16> {
281 match requested {
282 // FMT-9f: new files are v9; the `MNEMO_DEV_FORMAT` switch is gone.
283 0 => Ok(crate::format::DEFAULT_NEW_VERSION),
284 VERSION_V9 => Ok(requested),
285 // The v8 sunset (v0.13): v8 files open read-only and compact to v9;
286 // no new one is created.
287 v => Err(MnemoError::Invalid(format!(
288 "format_version {v} is not supported for new files: new files are v9 (pass 0 or 9)"
289 ))),
290 }
291}
292
293/// Catalog entry: maps a memory ID to its page run on disk.
294///
295/// **Schema (v5+):** carries `accessed_at` and `access_count` so `recall`
296/// can update access stats by mutating the catalog entry rather than
297/// rewriting the entire record (Phase 2.1 of the improvement plan). The
298/// values in a memory's serialized record body become a stale snapshot
299/// after the first recall; [`Mnemo::read_memory`] re-populates them from
300/// the catalog so consumers always see the live values.
301#[derive(Serialize, Deserialize, Clone, Debug)]
302struct CatalogEntry {
303 /// ULID as a raw `u128`.
304 id: u128,
305 start_page: u64,
306 page_count: u32,
307 /// Exact serialized byte length of the record.
308 len: u32,
309 deleted: bool,
310 /// Unix seconds of the most recent recall hit (or write, for fresh entries).
311 accessed_at: i64,
312 /// How many times `recall` has surfaced this memory.
313 access_count: u32,
314 /// FMT-9: the generation the record's pages were sealed under (v9 page
315 /// AAD). `0`, and left out of the encoding, on v8 files, so a v8
316 /// catalog is byte-for-byte what v0.10 writes.
317 #[serde(default, skip_serializing_if = "crate::format::is_zero_u64")]
318 gen: u64,
319 /// FMT-9b: the v9 row's tags, scoring inputs and vector location
320 /// (`None`, and left out of the encoding, on v8 files).
321 #[serde(default, skip_serializing_if = "Option::is_none")]
322 v9: Option<RowV9>,
323}
324
325/// FMT-9b: what a v9 catalog row carries beyond the v8 fields, so recall
326/// can filter and score a candidate without decrypting its record, and a
327/// cold handle can select records by kind or thread.
328#[derive(Serialize, Deserialize, Clone, Copy, Debug, Default)]
329struct RowV9 {
330 /// The vector column entry: page (`0` while it's waiting for the next
331 /// flush), slot, and the page's generation.
332 vec_page: u64,
333 vec_slot: u16,
334 vec_gen: u64,
335 /// FMT-9b.2: the record's slot on its packed page (`start_page`), or
336 /// [`REC_SPAN`] when the record owns a run of whole pages.
337 rec_slot: u16,
338 /// `ROW_KIND_*`.
339 kind: u8,
340 /// `ROW_SUPERSEDED | ROW_HAS_TTL | ROW_SHARED`.
341 flags: u8,
342 /// [`memory_type_code`].
343 memory_type: u8,
344 /// FNV-1a of `_mnemo.thread`, of the 12.1 key slot, and of `agent_id`
345 /// (`0` = none). A collision only lets a non-match through to the
346 /// decrypted check.
347 thread_hash: u32,
348 key_hash: u32,
349 agent_hash: u32,
350 created_at: i64,
351 importance: f32,
352 /// `created_at + ttl_secs` when `ROW_HAS_TTL` is set.
353 expires_at: i64,
354 /// FMT-9d: the row's BM25 document id (`0`: not keyword-indexed) and
355 /// its token count. Live rows define which documents are live, and
356 /// N and avgdl, so segments never hold lengths or tombstones.
357 doc: u32,
358 doc_len: u16,
359}
360
361/// FMT-9e: by default a thread turn is written plain and compressed once
362/// it's a week old (by its `created_at`), so a live session's recent turns
363/// read at full speed and archived history is small.
364pub const DEFAULT_TURN_COMPRESS_AFTER_SECS: i64 = 7 * 24 * 3600;
365/// Frozen v4 catalog entry shape, used only by the v4→v5 migration path in
366/// [`Mnemo::open`]. Its layout matches the catalog encoding written by
367/// pre-v5 builds (5 positional fields). Do not change this struct.
368#[derive(Serialize, Deserialize, Clone, Debug)]
369struct CatalogEntryV4 {
370 id: u128,
371 start_page: u64,
372 page_count: u32,
373 len: u32,
374 deleted: bool,
375}
376
377/// Returned by [`Mnemo::prepare_for_flush`] and consumed by [`Mnemo::flush`].
378/// Carries the already-serialized control-plane bytes through the prelude
379/// so we don't re-serialize after the lease has been persisted.
380struct FlushPrelude {
381 cat_bytes: Option<Vec<u8>>,
382 /// FMT-9c.2: `cat_bytes` is a delta run, not a base catalog.
383 cat_delta: bool,
384 /// FMT-9c.3: `idx_bytes` is an index delta run, not a whole index.
385 idx_delta: bool,
386 /// FMT-9d: this commit's BM25 segment, and the segments it replaces.
387 bm25_seg: Option<(Vec<u8>, Option<usize>)>,
388 idx_bytes: Option<Vec<u8>>,
389 /// Serialized result-cache directory bytes when `dirty_cache` is
390 /// set (Phase 10.1). `None` if the cache didn't change since the
391 /// last flush — in that case the header keeps its current
392 /// `cache_start` / `cache_pages` / `cache_len` and no fresh cache
393 /// run is written.
394 cache_bytes: Option<Vec<u8>>,
395}
396
397/// One entry in the append-only snapshot manifest. Because record, catalog,
398/// and index pages are only ever *appended*, the runs a past flush wrote are
399/// still on disk; a `Snapshot` is the set of pointers needed to reconstruct
400/// the database exactly as that flush left it.
401#[derive(Serialize, Deserialize, Clone, Debug)]
402struct Snapshot {
403 txn_id: u64,
404 created_at: i64,
405 catalog_start: u64,
406 catalog_pages: u64,
407 catalog_len: u64,
408 index_start: u64,
409 index_pages: u64,
410 index_len: u64,
411 memory_count: u64,
412 /// FMT-9: generations of the catalog and index runs (v9 only; `None`,
413 /// and left out of the encoding, on v8 files).
414 #[serde(default, skip_serializing_if = "Option::is_none")]
415 gens: Option<SnapshotGens>,
416}
417
418/// FMT-9: the generations a v9 snapshot's runs were sealed under.
419#[derive(Serialize, Deserialize, Clone, Copy, Debug, Default)]
420struct SnapshotGens {
421 catalog: u64,
422 index: u64,
423 /// FMT-9c.2: the catalog's delta-chain head at this snapshot.
424 #[serde(default)]
425 delta: (u64, u64, u64, u64),
426 /// FMT-9c.3: the index's delta-chain head at this snapshot.
427 #[serde(default)]
428 index_delta: (u64, u64, u64, u64),
429 /// FMT-9d: the BM25 manifest at this snapshot.
430 #[serde(default)]
431 bm25: (u64, u64, u64, u64),
432}
433
434impl Snapshot {
435 fn catalog_gen(&self) -> u64 {
436 self.gens.map_or(0, |g| g.catalog)
437 }
438 fn index_gen(&self) -> u64 {
439 self.gens.map_or(0, |g| g.index)
440 }
441 fn delta_head(&self) -> RunPtr {
442 self.gens.map_or(RunPtr::default(), |g| tuple_ptr(g.delta))
443 }
444 fn index_delta_head(&self) -> RunPtr {
445 self.gens
446 .map_or(RunPtr::default(), |g| tuple_ptr(g.index_delta))
447 }
448}
449
450/// A restorable point in a database's history — see [`Mnemo::snapshots`].
451#[derive(Clone, Copy, Debug)]
452pub struct SnapshotInfo {
453 /// Id of the transaction that produced this snapshot (monotonic from 1).
454 pub txn_id: u64,
455 /// When the snapshot was committed (unix seconds).
456 pub created_at: i64,
457 /// Live memory count captured in the snapshot.
458 pub memory_count: u64,
459}
460
461/// What [`Mnemo::stats`] and [`Mnemo::stats_summary`] report about restore
462/// points (INSP-1, RB-001.25).
463pub(super) struct RestorePoints {
464 pub(super) retained: usize,
465 pub(super) oldest_at: Option<i64>,
466 pub(super) pinned_bytes: u64,
467}
468
469/// A retrieval request for [`Mnemo::recall`].
470#[derive(Clone, Debug)]
471pub struct RecallRequest {
472 /// Query embedding.
473 pub query: Vec<f32>,
474 /// Maximum results to return.
475 pub top_k: usize,
476 /// Restrict to these memory types (`None` = all types).
477 pub memory_types: Option<Vec<MemoryType>>,
478 /// Restrict to a single agent's view: its own memories plus shared ones.
479 pub agent_id: Option<String>,
480 /// Similarity metric.
481 pub metric: Metric,
482 /// Multi-signal score weights.
483 pub weights: ScoreWeights,
484 /// Index override: partitions to probe (`None` = index default). Ignored
485 /// when no ANN index is present.
486 pub n_probe: Option<usize>,
487 /// Index override: candidates to rerank (`None` = index default). Ignored
488 /// when no ANN index is present.
489 pub n_rerank: Option<usize>,
490 /// Whether to update each result's `accessed_at` and `access_count` in
491 /// the catalog. Defaults to `true`, matching pre-v5 behavior.
492 ///
493 /// Set to `false` for **fully read-only recall** — useful for batch
494 /// scoring, dry-run scoring, or recalls run by introspection tooling
495 /// (e.g. `mnemo about` consumers) where you don't want the score's own
496 /// observation to perturb the database. With `false`, recall does not
497 /// dirty the catalog and the next `flush` is a no-op.
498 pub track_access: bool,
499 /// Phase 12.6 pre-scoring filter: `session_id` + AND-ed metadata
500 /// conditions applied before the candidate is scored. Default is
501 /// [`Filter::default`] (no constraints). When the ANN index is on
502 /// and this filter is selective enough to starve the shortlist
503 /// below `top_k`, Phase 12.1's widen-then-exact-scan logic kicks
504 /// in so the result still carries up to `top_k` correct hits.
505 pub filter: Filter,
506 /// 13a (13.5): the visibility levels recall returns: public and team by
507 /// default; private only when asked ([`RecallRequest::visibility`]).
508 pub visibility: crate::memory::VisibilitySet,
509 /// Phase 12.8 post-scoring budget: stop adding hits once the UTF-8
510 /// bytes of their `content` fields (not the serialized record
511 /// body) would exceed this many bytes. `None` (default) means no
512 /// budget and every hit through `top_k` is returned. If the top-1
513 /// hit is already over budget it is still returned — callers that
514 /// need an absolute cap should also pre-size `top_k`.
515 ///
516 /// Counts only the UTF-8 `content` bytes so the number is stable
517 /// across the Rust / Python / CLI / MCP surfaces; the JSON or
518 /// MessagePack envelope is excluded. Roughly 4 bytes per BPE
519 /// token for English text — a 4 KB budget comfortably holds ~1K
520 /// tokens.
521 ///
522 /// Observable through [`RecallResponse::truncated`]
523 /// ([`Mnemo::recall_budgeted`]): `true` iff at least one scored
524 /// hit was dropped by this budget, `false` otherwise.
525 pub max_bytes: Option<usize>,
526 /// FMT-9d (Phase 2, v9 files): **identifiers or exact terms to match**
527 /// (a version, a PR number, a function name) for hybrid recall. BM25
528 /// scores, divided by this query's best, are added to each candidate's
529 /// similarity with weight [`RecallRequest::keyword_weight`] before the
530 /// usual scoring. With an empty `query` vector
531 /// ([`RecallRequest::keywords_only`]), recall is keyword-only.
532 ///
533 /// **Since v0.15 (12.3 step 4), a recall by text is hybrid by default:**
534 /// with `keywords` `None`, the request's [`RecallRequest::text`] is
535 /// used for BM25, so give keywords only for identifiers the question
536 /// doesn't contain. Keywords given replace the text. **Empty keywords
537 /// (`.keywords("")`) opt out:** vector-only recall. `None` on a request
538 /// with no text is plain vector recall.
539 pub keywords: Option<String>,
540 /// 12.3 step 4 (v0.15): the query text a recall by text was built from
541 /// ([`RecallRequest::from_text`] sets it). With `keywords` `None` it's
542 /// also scored with BM25 on a v9 file (hybrid by default); a v8 file,
543 /// which has no keyword index, ignores it.
544 pub text: Option<String>,
545 /// FMT-9d: weight ε of the normalised BM25 score in hybrid recall
546 /// (default [`DEFAULT_KEYWORD_WEIGHT`]), for keywords given.
547 pub keyword_weight: f32,
548 /// 12.3 step 4 (v0.15): the weight when BM25 scores the request's
549 /// [`RecallRequest::text`] (no keywords given); default
550 /// [`DEFAULT_TEXT_KEYWORD_WEIGHT`].
551 pub text_keyword_weight: f32,
552 /// FMT-9d: BM25 term-frequency saturation `k1` (default 1.2).
553 pub bm25_k1: f32,
554 /// FMT-9d: BM25 length normalisation `b` (default 0.75).
555 pub bm25_b: f32,
556}
557
558impl RecallRequest {
559 /// A request with sensible defaults for a query embedding.
560 pub fn new(query: Vec<f32>) -> Self {
561 Self {
562 query,
563 top_k: 10,
564 memory_types: None,
565 agent_id: None,
566 metric: Metric::Cosine,
567 weights: ScoreWeights::default(),
568 n_probe: None,
569 n_rerank: None,
570 track_access: true,
571 filter: Filter::default(),
572 visibility: crate::memory::VisibilitySet::DEFAULT,
573 max_bytes: None,
574 keywords: None,
575 text: None,
576 keyword_weight: DEFAULT_KEYWORD_WEIGHT,
577 text_keyword_weight: DEFAULT_TEXT_KEYWORD_WEIGHT,
578 bm25_k1: 1.2,
579 bm25_b: 0.75,
580 }
581 }
582
583 /// 13a (13.5): the visibility levels to return (private records only
584 /// when the set includes them).
585 pub fn visibility(mut self, v: crate::memory::VisibilitySet) -> Self {
586 self.visibility = v;
587 self
588 }
589
590 /// FMT-9d (v9 files): a keyword-only request: BM25 over `keywords`
591 /// (identifiers or exact terms), no query vector, so it works without
592 /// an embedder.
593 pub fn keywords_only(keywords: impl Into<String>) -> Self {
594 Self::new(Vec::new()).keywords(keywords)
595 }
596
597 /// FMT-9d (v9 files): add identifiers or exact terms to match (hybrid
598 /// recall). They replace the request's text for BM25; `""` opts out of
599 /// keyword scoring. See the `keywords` field.
600 pub fn keywords(mut self, text: impl Into<String>) -> Self {
601 self.keywords = Some(text.into());
602 self
603 }
604
605 /// 12.3 step 4: the weight of the text's keyword score when no keywords
606 /// are given (see [`RecallRequest::text_keyword_weight`]).
607 pub fn text_keyword_weight(mut self, w: f32) -> Self {
608 self.text_keyword_weight = w;
609 self
610 }
611
612 /// 12.3 step 4: the weight of this request's keyword score: the text's
613 /// when it's the text that's scored, the keywords' otherwise.
614 pub(crate) fn effective_keyword_weight(&self) -> f32 {
615 if self.keywords.is_none() {
616 self.text_keyword_weight
617 } else {
618 self.keyword_weight
619 }
620 }
621
622 /// 12.3 step 4: the text BM25 scores for this request: the keywords
623 /// given (none if they're empty), or else the query text on a v9 file.
624 pub(crate) fn keywords_for(&self, v9: bool) -> Option<&str> {
625 match self.keywords.as_deref() {
626 Some(k) if k.trim().is_empty() => None,
627 Some(k) => Some(k),
628 None if v9 => self.text.as_deref().filter(|t| !t.trim().is_empty()),
629 None => None,
630 }
631 }
632
633 /// FMT-9d: the weight ε of the keyword score in hybrid recall.
634 pub fn keyword_weight(mut self, w: f32) -> Self {
635 self.keyword_weight = w;
636 self
637 }
638
639 /// Phase 12.6: attach a pre-scoring filter. See [`Filter`] and
640 /// [`FilterTerm`] for term shapes.
641 pub fn filter(mut self, filter: Filter) -> Self {
642 self.filter = filter;
643 self
644 }
645
646 /// Phase 12.8: cap the total UTF-8 bytes of returned `content`
647 /// fields. See [`RecallRequest::max_bytes`].
648 pub fn max_bytes(mut self, n: usize) -> Self {
649 self.max_bytes = Some(n);
650 self
651 }
652
653 /// Build a request by embedding `query_text` with the supplied
654 /// embedder (Phase 3). Since v0.15 it embeds with
655 /// [`Embedder::embed_query`] and keeps the text in
656 /// [`RecallRequest::text`], so the recall is hybrid by default (see
657 /// [`RecallRequest::keywords`]; `.keywords("")` opts out). Otherwise the
658 /// same defaults as `new`. Fails if the embedder returns an empty vector
659 /// list or an embedding failure.
660 pub fn from_text(embedder: &dyn Embedder, query_text: &str) -> Result<Self> {
661 let mut vectors = embedder.embed_query(&[query_text])?;
662 let v = vectors.pop().ok_or_else(|| {
663 MnemoError::Invalid("embedder returned no vector for the query text".into())
664 })?;
665 Ok(Self::new(v).text(query_text))
666 }
667
668 /// 12.3 step 4: set the query text (see [`RecallRequest::text`]), for
669 /// a caller that embedded it itself.
670 pub fn text(mut self, text: impl Into<String>) -> Self {
671 self.text = Some(text.into());
672 self
673 }
674 /// Set the result cap.
675 pub fn top_k(mut self, k: usize) -> Self {
676 self.top_k = k;
677 self
678 }
679 /// Restrict to specific memory types.
680 pub fn types(mut self, t: Vec<MemoryType>) -> Self {
681 self.memory_types = Some(t);
682 self
683 }
684 /// Restrict to one agent's view.
685 pub fn agent(mut self, agent_id: impl Into<String>) -> Self {
686 self.agent_id = Some(agent_id.into());
687 self
688 }
689 /// Set the similarity metric (default: cosine).
690 pub fn metric(mut self, metric: Metric) -> Self {
691 self.metric = metric;
692 self
693 }
694 /// Replace the multi-signal score weights.
695 pub fn weights(mut self, weights: ScoreWeights) -> Self {
696 self.weights = weights;
697 self
698 }
699 /// Override the number of IVF partitions probed (accuracy/speed dial).
700 pub fn n_probe(mut self, n: usize) -> Self {
701 self.n_probe = Some(n);
702 self
703 }
704 /// Override the number of candidates reranked exactly (accuracy dial).
705 pub fn n_rerank(mut self, n: usize) -> Self {
706 self.n_rerank = Some(n);
707 self
708 }
709 /// Toggle whether recall updates `accessed_at` / `access_count` on
710 /// the returned memories' catalog entries (default `true`). Pass
711 /// `false` for a fully read-only recall — see the field doc.
712 pub fn track_access(mut self, track: bool) -> Self {
713 self.track_access = track;
714 self
715 }
716}
717
718/// One scored result from [`Mnemo::recall`].
719#[derive(Clone, Debug)]
720pub struct RecallResult {
721 /// The retrieved memory.
722 pub memory: Memory,
723 /// Combined multi-signal score.
724 pub score: f32,
725 /// The bare similarity component (before other signals).
726 pub similarity: f32,
727}
728
729/// Phase 12.8 budgeted-recall envelope returned by
730/// [`Mnemo::recall_budgeted`]. Wraps the `Vec<RecallResult>` the
731/// legacy [`Mnemo::recall`] returns so callers can tell whether
732/// [`RecallRequest::max_bytes`] cut anything.
733#[derive(Clone, Debug)]
734pub struct RecallResponse {
735 /// Scored hits, best first, truncated to at most `top_k` and then
736 /// to fit the `max_bytes` budget if set.
737 pub hits: Vec<RecallResult>,
738 /// `true` iff the `max_bytes` budget dropped at least one hit
739 /// that would otherwise have been returned (i.e. the result is
740 /// shorter than what `top_k` alone would produce).
741 pub truncated: bool,
742 /// Phase 12.1 + 12.6 observability: `true` iff recall fell
743 /// through to an exact scan (ANN shortlist too narrow after
744 /// filtering, and the widen pass also came up short, so the
745 /// index was skipped for one full-corpus pass). Monitoring on
746 /// this flag is how callers detect selective-filter regressions
747 /// or supersession chains outgrowing the ANN shortlist.
748 pub used_exact_scan: bool,
749}
750
751/// Where a file's bytes and an open handle's heap go (P-0 device budget,
752/// `examples/device_budget.rs`). **Unstable diagnostic**: fields may
753/// change in any release. File figures come from the header and the
754/// catalog. RAM figures are buffer capacities, with hash maps estimated
755/// from their capacity; they cover the engine's main structures, not
756/// every allocation, and not allocator overhead.
757#[doc(hidden)]
758#[derive(Clone, Debug, Default, serde::Serialize)]
759#[cfg(feature = "dev-hooks")]
760pub struct BudgetReport {
761 pub file_bytes: u64,
762 pub page_size: u64,
763 /// Page 0.
764 pub header_bytes: u64,
765 /// Whole pages holding live records (superseded versions included).
766 pub record_page_bytes: u64,
767 /// Serialized bytes of those records; the rest of their pages is
768 /// slack.
769 pub record_payload_bytes: u64,
770 /// Pages still holding soft-deleted records (reclaimed by compact).
771 pub deleted_record_page_bytes: u64,
772 pub catalog_run_bytes: u64,
773 pub catalog_payload_bytes: u64,
774 pub index_run_bytes: u64,
775 pub index_payload_bytes: u64,
776 /// First page of the persisted index run (0 = none).
777 pub index_start_page: u64,
778 pub wal_bytes: u64,
779 pub snapshot_manifest_bytes: u64,
780 pub cache_dir_run_bytes: u64,
781 /// Everything else: superseded catalog, index and directory runs from
782 /// earlier flushes, and freed pages, until the next compact.
783 pub unreferenced_bytes: u64,
784 pub live_records: u64,
785 pub ram_catalog: u64,
786 pub ram_id_map: u64,
787 pub ram_ann_centroids: u64,
788 pub ram_ann_codebook: u64,
789 pub ram_ann_posting_ids: u64,
790 pub ram_ann_posting_codes: u64,
791 pub ram_ann_assignment: u64,
792 pub ram_cache_dir: u64,
793 pub ram_key_caches: u64,
794 /// FMT-9b: the resident fp16 vector column as allocated (rows not yet
795 /// touched by recall cost no physical memory).
796 pub ram_vector_column: u64,
797 /// Sum of the `ram_*` fields.
798 pub ram_total: u64,
799}
800
801/// Summary statistics for a database.
802#[derive(Clone, Debug)]
803pub struct Stats {
804 /// Live (non-deleted) memory count.
805 pub memories: usize,
806 /// Tombstoned entries awaiting compaction.
807 pub deleted: usize,
808 /// Embedding dimensionality.
809 pub dimensions: usize,
810 /// File size in bytes.
811 pub file_bytes: u64,
812 /// Distinct agent IDs present.
813 pub agents: Vec<String>,
814 /// Whether pages are encrypted (always true in v1).
815 pub encrypted: bool,
816 /// Creation time (unix seconds).
817 pub created_at: i64,
818 /// ANN index shape, if an index has been built.
819 pub index: Option<IndexInfo>,
820 /// Current size of the write-ahead log region, in 8 KiB pages.
821 pub wal_pages: u64,
822 /// Fsyncs issued by the most recent successful [`Mnemo::flush`] on
823 /// this handle. `0` on a fresh handle before any flush, or when
824 /// `flush()` short-circuited (nothing dirty). Phase 6 instrumentation
825 /// for the Phase 11.1 write-latency work — the planner wants this
826 /// before deciding whether cutting fsync count is a viable
827 /// durability-preserving fix.
828 pub fsyncs_last_flush: u32,
829 /// FMT-9f: the file's on-disk format (`8` or `9`). Since v0.13 a v8 file
830 /// opens read-only; `mnemo compact` converts it to v9.
831 pub format_version: u16,
832 /// MOBILE-2d: how recall scores vectors on this handle right now.
833 pub recall_mode: RecallMode,
834 /// INSP-1 (RB-001.25): the snapshots retained right now, each a
835 /// restore point for [`Mnemo::restore_to`]. Under churn the pin cap can
836 /// keep fewer than the policy's 16-and-24-hours, so this, not the
837 /// policy, is what can be restored.
838 pub snapshots_retained: usize,
839 /// When the oldest retained snapshot was committed (unix seconds):
840 /// the furthest back a restore reaches. `None` with no snapshot.
841 pub oldest_restorable_at: Option<i64>,
842 /// The bytes the retained snapshots keep from reuse (pages freed after
843 /// the oldest one). `0` on v8, which never reuses pages.
844 pub snapshot_pinned_bytes: u64,
845 /// TOK-1 (v0.15): the keyword index's tokenizer: `V2` for a file v0.15
846 /// created or rebuilt, `V1` for an older file until a compaction or
847 /// `reembed` rebuilds it, `Mixed` after a build before v0.15 wrote to
848 /// a tokenizer 2 index. `None` on v8, which has no keyword index.
849 pub keyword_tokenizer: Option<crate::KeywordTokenizer>,
850}
851
852/// [`Mnemo::stats_summary`]: the cheap statistics, read from the catalog and
853/// the header (v0.12.2, RB-002.3).
854#[derive(Clone, Debug, PartialEq, Eq)]
855pub struct StatsSummary {
856 /// Live (non-deleted) memory count.
857 pub memories: usize,
858 /// Tombstoned entries awaiting compaction.
859 pub deleted: usize,
860 /// Embedding dimensionality.
861 pub dimensions: usize,
862 /// File size in bytes.
863 pub file_bytes: u64,
864 /// The on-disk format, `8` or `9`.
865 pub format_version: u16,
866 /// Current size of the write-ahead log region, in 8 KiB pages.
867 pub wal_pages: u64,
868 /// Whether an IVF+PQ index exists (loaded, or persisted in the file).
869 pub has_index: bool,
870 /// How recall scores vectors right now.
871 pub recall_mode: RecallMode,
872 /// The bytes the v9 vector column needs to stay in memory: the live
873 /// facts × (`2 × dimensions + 4`). Recall is `ExactResident` while the
874 /// vector cache budget is at least this; `0` on v8, which has no column.
875 pub vector_column_bytes: u64,
876 /// INSP-1 (RB-001.25): the snapshots retained right now, each a
877 /// restore point for [`Mnemo::restore_to`]. Under churn the pin cap can
878 /// keep fewer than the policy's 16-and-24-hours, so this, not the
879 /// policy, is what can be restored.
880 pub snapshots_retained: usize,
881 /// When the oldest retained snapshot was committed (unix seconds):
882 /// the furthest back a restore reaches. `None` with no snapshot.
883 pub oldest_restorable_at: Option<i64>,
884 /// The bytes the retained snapshots keep from reuse (pages freed after
885 /// the oldest one). `0` on v8, which never reuses pages.
886 pub snapshot_pinned_bytes: u64,
887}
888
889/// MOBILE-2d: how recall scores vectors, from [`Mnemo::stats`].
890#[derive(Clone, Copy, Debug, PartialEq, Eq)]
891pub enum RecallMode {
892 /// An exact scan over the fp16 vector column held in memory: a v9 file
893 /// whose live facts fit [`Mnemo::set_vector_cache_budget`]. The fast
894 /// path.
895 ExactResident,
896 /// An exact scan that reads vectors through the page cache: a v9 file
897 /// whose live facts are over the budget, or a v8 file. Slower, and
898 /// slower still as the file grows; raise the budget or build an index.
899 ExactPaged,
900 /// An IVF+PQ index shortlists the candidates
901 /// ([`Mnemo::build_index`]).
902 Indexed,
903}
904
905impl RecallMode {
906 /// `exact_resident`, `exact_paged` or `indexed`.
907 pub fn as_str(&self) -> &'static str {
908 match self {
909 RecallMode::ExactResident => "exact_resident",
910 RecallMode::ExactPaged => "exact_paged",
911 RecallMode::Indexed => "indexed",
912 }
913 }
914}
915
916/// The error from [`Mnemo::try_close`]: the final flush failed, and the
917/// handle comes back with it, still open and still holding every write that
918/// wasn't flushed.
919pub struct CloseError {
920 pub(super) handle: Box<Mnemo>,
921 pub(super) error: MnemoError,
922}
923
924impl CloseError {
925 /// Why the flush failed.
926 pub fn error(&self) -> &MnemoError {
927 &self.error
928 }
929
930 /// The handle, to retry [`Mnemo::flush`] or [`Mnemo::try_close`].
931 pub fn into_inner(self) -> Mnemo {
932 *self.handle
933 }
934
935 /// The handle and the error.
936 pub fn into_parts(self) -> (Mnemo, MnemoError) {
937 (*self.handle, self.error)
938 }
939}
940
941impl std::fmt::Debug for CloseError {
942 fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
943 f.debug_struct("CloseError")
944 .field("error", &self.error)
945 .finish_non_exhaustive()
946 }
947}
948
949impl std::fmt::Display for CloseError {
950 fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
951 write!(f, "close failed (the handle is kept): {}", self.error)
952 }
953}
954
955impl std::error::Error for CloseError {
956 fn source(&self) -> Option<&(dyn std::error::Error + 'static)> {
957 Some(&self.error)
958 }
959}
960
961/// Result of a [`Mnemo::compact_file`] run.
962#[derive(Clone, Copy, Debug, Default)]
963pub struct CompactReport {
964 /// Live memories before compaction.
965 pub before: usize,
966 /// Live memories after compaction (expired ones dropped).
967 pub after: usize,
968 /// Live, non-expired cache entries before compaction. FIX-1
969 /// R-001.1: compaction now preserves the result cache instead of
970 /// silently dropping it.
971 pub cache_before: usize,
972 /// Live, non-expired cache entries after compaction (tombstoned
973 /// and TTL-expired ones dropped).
974 pub cache_after: usize,
975 /// Legacy onboarding or manifest memories that held an all-zero vector
976 /// (from before the zero-vector rule) and were given the sentinel unit
977 /// vector a current manifest has.
978 pub zero_vector_manifests: usize,
979 /// Other memories with an all-zero vector, carried unchanged: their
980 /// content and recall behaviour are the same as before.
981 pub zero_vector_memories: usize,
982}
983
984/// What [`Mnemo::reembed_file`] did (12.3 step 4, v0.15).
985#[derive(Clone, Debug, PartialEq, Eq)]
986pub struct ReembedReport {
987 /// Live memories embedded again: facts, superseded versions, essence
988 /// items and thread turns (expired ones are dropped, as by compaction).
989 pub memories: usize,
990 /// Manifests kept on the sentinel vector (resized to the new width).
991 pub manifests: usize,
992 /// Thread links (`_mnemo.kind = "thread_link"`) kept on the sentinel
993 /// vector (resized to the new width), R-008.3.
994 pub thread_links: usize,
995 /// Result-cache entries copied (the engine's own aren't counted).
996 pub cache_entries: usize,
997 /// Semantic cache entries dropped because their vector has the old
998 /// width (only when the width changed).
999 pub cache_dropped: usize,
1000 /// The identity the file recorded before, if any.
1001 pub from: Option<crate::EmbedderIdentity>,
1002 /// The identity it records now.
1003 pub to: crate::EmbedderIdentity,
1004}
1005
1006/// What [`Mnemo::verify_report`] checked.
1007#[derive(Clone, Debug, PartialEq, Eq)]
1008pub struct VerifyReport {
1009 /// Records decrypted and decoded: live memories and cache entries (the
1010 /// engine's own entries aren't counted).
1011 pub records: usize,
1012 /// Live memories whose vector is all zeros (from before the zero-vector
1013 /// rule). They score 0.0 on every vector query.
1014 pub zero_vector_memories: usize,
1015 /// TOK-1: the keyword index's tokenizer (see
1016 /// [`Stats::keyword_tokenizer`]). `Mixed` is worth a compaction.
1017 pub keyword_tokenizer: Option<crate::KeywordTokenizer>,
1018}
1019
1020/// An encrypted, single-file agent memory database.
1021pub struct Mnemo {
1022 /// 12.4: the lock layer's note for a handle opened through it for a
1023 /// long operation ([`Mnemo::open_exclusive`], compaction): its
1024 /// `lock-info` line, removed when the handle drops. Declared first so
1025 /// it drops before `pager`, whose file's unlock comes after it.
1026 lock_note: Option<crate::shared::Held>,
1027 /// 12.3: what this handle has to tell its owner once, such as a
1028 /// passphrase on raw-key KDF parameters ([`Mnemo::notices`]).
1029 notices: Vec<String>,
1030 pager: Pager,
1031 header: Header,
1032 catalog: Vec<CatalogEntry>,
1033 index: IdIndex,
1034 /// The path this handle opened, made absolute at open (so a later
1035 /// change of the process's directory doesn't change what it names).
1036 /// `ensure_not_replaced` checks that it still names the open file.
1037 path: PathBuf,
1038 dimensions: usize,
1039 kdf: KdfParams,
1040 /// Set whenever the catalog changes; drives whether `flush` rewrites it.
1041 dirty_catalog: bool,
1042 /// Optional IVF+PQ approximate-nearest-neighbour index.
1043 ann: Option<IvfPqIndex>,
1044 /// Set whenever the in-memory ANN index differs from the persisted one.
1045 /// PERF-1: `flush` persists it only per [`IndexPersist`]'s policy.
1046 dirty_index: bool,
1047 /// PERF-1: index persistence policy state.
1048 index_persist: IndexPersist,
1049 /// PERF-1 step 3: the header points at a persisted ANN index that
1050 /// hasn't been loaded yet. It loads on first use (`ensure_ann`).
1051 ann_unloaded: bool,
1052 /// Append-only manifest of committed snapshots, oldest first. Capped
1053 /// at [`Mnemo::max_snapshots`] entries on every flush — the oldest
1054 /// entries are pruned first.
1055 manifest: Vec<Snapshot>,
1056 /// Maximum manifest entries to retain across flushes. `0` disables
1057 /// the cap (retain forever). Defaults to [`DEFAULT_MAX_SNAPSHOTS`].
1058 max_snapshots: usize,
1059 /// If `true`, this handle was opened via [`Mnemo::open_read_only`] and
1060 /// every mutating API returns [`MnemoError::ReadOnly`]. The underlying
1061 /// file was opened without write permission and holds an OS shared
1062 /// lock; the OS would reject a write() anyway, but gating in Rust
1063 /// gives callers a clean, typed error.
1064 read_only: bool,
1065 /// Result-cache directory (Phase 10.1). See [`crate::result_cache`].
1066 /// Loaded from `header.cache_start` on open; persisted on `flush`
1067 /// via the same seal_run pipeline as catalog/index/manifest.
1068 cache_dir: Vec<CacheDirectoryEntry>,
1069 /// In-memory index over `cache_dir` for O(1) `(namespace, key_hash)`
1070 /// lookups. Rebuilt on open + kept in sync on `cache_put`/`delete`.
1071 cache_idx: DirectoryIndex,
1072 /// Set whenever `cache_dir` changes; drives whether `flush` rewrites
1073 /// the cache directory run.
1074 dirty_cache: bool,
1075 /// Per-namespace eviction budgets. Missing entries fall back to
1076 /// [`CacheBudget::default`]. Not persisted (per-handle policy).
1077 cache_budgets: std::collections::HashMap<String, CacheBudget>,
1078 /// Flush policy for cache mutations. Memory writes ignore this and
1079 /// always behave as `Strict`.
1080 cache_flush_policy: CacheFlushPolicy,
1081 /// Bookkeeping for `CacheFlushPolicy::Batched` — pending mutation
1082 /// count and first-dirty timestamp.
1083 cache_batch: BatchState,
1084 /// In-memory hit/miss/eviction counters (reset on reopen).
1085 cache_counters: CounterState,
1086 /// Optional handle-scoped embedder (Phase 3, PR 10a). Registered
1087 /// via [`Mnemo::set_embedder`]; consulted by [`Mnemo::remember_text`]
1088 /// and [`RecallRequest::from_text`]. `None` means text-side APIs
1089 /// return [`MnemoError::Invalid`] with a "no embedder" message.
1090 /// Not persisted — the handle is re-registered after every open.
1091 embedder: Option<std::sync::Arc<dyn Embedder>>,
1092 /// Fsyncs issued by the most recent `flush()` call. Snapshotted
1093 /// from the thread-local fsync counter in `lib.rs::fsync_counter`:
1094 /// `flush()` reads it on entry, does its work (each `sync_all` the
1095 /// pager / wal / recovery path makes bumps the counter), and
1096 /// stores the delta here. `0` on a fresh handle before any flush,
1097 /// or on a read-only handle where `flush()` short-circuits.
1098 /// Exposed through [`Stats::fsyncs_last_flush`] for diagnostics —
1099 /// the Phase 6 planner wants this before deciding whether cutting
1100 /// fsync count is a viable write-latency win.
1101 fsyncs_last_flush: u32,
1102 /// FIX-1 R-001.8 part 2: lazy cache of the distinct agent IDs
1103 /// across every live memory, built once on first `stats()` call
1104 /// per handle. `remember()` adds to it in O(1); `delete()` can't
1105 /// cheaply tell whether the removed memory was the last of its
1106 /// agent (the agent lives in the encrypted body, not the catalog
1107 /// — pre-v9 format), so it just invalidates the set and the next
1108 /// `stats()` call rebuilds. Pre-FIX-1 every `stats()` call
1109 /// decrypted every live memory to collect agents, turning an
1110 /// otherwise-cheap diagnostic into an O(N) decrypt.
1111 agent_set: Option<std::collections::BTreeSet<String>>,
1112 /// Phase 12.1 lazy caches, built together on the first keyed op this
1113 /// handle sees (`remember_keyed`, `update` that touches `_mnemo.key`,
1114 /// or any `history(...)` call). Both are maintained incrementally
1115 /// after that. `None` → not built yet; `Some(…)` → authoritative.
1116 ///
1117 /// * `key_index`: `(scope_owner, key) → ULID bytes` of the **live**
1118 /// record for each keyed slot. `scope_owner` is `Some(agent_id)`
1119 /// for a `Private` scope slot and `None` for a `Shared` one, so a
1120 /// private write never collides with a shared one.
1121 /// * `superseded`: the ULID bytes of every record that has been
1122 /// superseded (`_mnemo.superseded_by` set on it). `recall` and
1123 /// `about` skip these before any decrypt.
1124 ///
1125 /// Building both costs one scan over live records (~1s at 100K with
1126 /// hardware AES) and happens at most once per handle. A format bump
1127 /// to v9 moves the key into the catalog row so the scan goes away.
1128 key_index: Option<std::collections::HashMap<KeySlot, u128>>,
1129 superseded: Option<std::collections::HashSet<u128>>,
1130 /// R-004.2: thread layers + onboarding ids, built by the same lazy scan.
1131 layers: Option<LayerCache>,
1132 /// REKEY-1 (RB-004.2): what the last successful rekey couldn't scrub
1133 /// once the new secret was in force (see [`Mnemo::rekey_cleanup_error`]).
1134 rekey_cleanup: Option<String>,
1135 /// 12.4 shared mode: access statistics from reads made under the
1136 /// shared lock (recall hits, cache hits), applied with this handle's
1137 /// next commit. `None` on every other handle.
1138 access_delta: Option<shared::AccessDelta>,
1139 /// 12.4 shared mode: the transaction id of the commit `flush` had
1140 /// reached the WAL with, when it fails after that point. A shared write
1141 /// uses it to tell a lost write from one that may be in the file.
1142 commit_point: Option<u64>,
1143 /// 12.4 shared mode (RB-007.5): a `flush` since the window started
1144 /// committed. Set by every committing flush; a shared write window
1145 /// clears it first, and then knows that part of the window is in the
1146 /// file when a later step fails.
1147 window_committed: bool,
1148 /// 12.4 shared mode: this handle is a shared handle's state, so a
1149 /// `rekey` inside a window is refused (`SharedMnemo::rekey` does it and
1150 /// updates the kept secret).
1151 shared: bool,
1152 /// FMT-9e: thread turns are compressed once older than this many
1153 /// seconds (`0`: at write). Per handle, not persisted.
1154 turn_compress_after: i64,
1155 /// 13a: stamped as `_mnemo.author` on each new record this handle
1156 /// writes ([`Mnemo::set_author`]).
1157 author: Option<String>,
1158 /// 13a (13.3): this handle's process holds the author's presence lock
1159 /// (a shared handle's [`SharedMnemo::hold_presence`]), so its claims
1160 /// may lapse early when it dies.
1161 presence_held: bool,
1162 /// 13a: seconds added to the clock coordination reads (tests only, via
1163 /// a dev-hooks setter).
1164 clock_offset: i64,
1165 /// RB-013.1: the ids of coordination records whose v9 row an older
1166 /// build re-tagged as a fact (its `compact` rewrites every row), found
1167 /// once per handle by decrypting the `episodic` fact rows.
1168 legacy_coordination: Option<std::collections::HashSet<u128>>,
1169 /// FMT-9 (a): the committed free list (v9; always empty on v8). Pages
1170 /// allocated from it since the last commit are already removed.
1171 free: FreeList,
1172 /// FMT-9 (a): runs this transaction stopped referencing (overwritten
1173 /// or deleted records, tombstoned cache entries, a relocated WAL
1174 /// region). The next commit adds them to the free list, freed at its
1175 /// own transaction id.
1176 pending_free: Vec<(u64, u64)>,
1177 /// FMT-9 (a): the free list changed (an allocation took pages from it)
1178 /// and must be rewritten by the next commit.
1179 free_dirty: bool,
1180 /// FMT-9 (b): the key-slot run (v9; empty until Phase 13.2 adds slots).
1181 key_slots: KeySlotRun,
1182 /// FMT-9 (b): `key_slots` changed since the last commit.
1183 dirty_keys: bool,
1184 /// FMT-9: this writable session has burned a transaction id with a
1185 /// durable header-only commit (`ensure_gen_lease`), so the generations
1186 /// its pages carry can't repeat ones a crashed earlier session used.
1187 gen_leased: bool,
1188 /// FMT-9b: the fp16 vector column (v9; unused on v8).
1189 col: VecColumn,
1190 /// FMT-9b.2: packed record pages (v9; unused on v8).
1191 recs: RecPages,
1192 /// FMT-9d: the BM25 keyword index (v9; unused on v8).
1193 bm25: Bm25State,
1194 /// FMT-9d: test hook: write this layout byte instead of the current
1195 /// one (to exercise the unknown-layout path).
1196 index_layout_override: Option<u8>,
1197 /// FMT-9d: during one recall, each keyword match's normalised BM25
1198 /// score, by id.
1199 kw: Option<std::collections::HashMap<u128, f32>>,
1200 /// FMT-9c.2: catalog rows changed since the last commit (v9 writes
1201 /// them as a delta run instead of the whole catalog).
1202 dirty_rows: std::collections::BTreeSet<usize>,
1203 /// FMT-9c.2: the delta runs on top of the base catalog, oldest first,
1204 /// and how many rows they hold in all.
1205 delta_chain: Vec<RunPtr>,
1206 delta_rows: usize,
1207 /// FMT-9c.3: the ANN index's delta runs on top of its base run, oldest
1208 /// first, and how many changes they hold in all.
1209 index_chain: Vec<RunPtr>,
1210 index_chain_ops: usize,
1211 /// FMT-9c.3: the next commit writes the whole index (after a build,
1212 /// rebuild or restore).
1213 index_rebase: bool,
1214 /// FMT-9c.3: this flush's index changes, when it writes them as a
1215 /// delta run.
1216 index_delta_ops: Option<Vec<IndexOp>>,
1217 /// FMT-9c.2: the next commit writes a whole base catalog (after a
1218 /// restore, or when the chain has grown too long).
1219 catalog_rewrite: bool,
1220 /// FMT-9c: the snapshot count was set explicitly (config or
1221 /// `set_max_snapshots`), which replaces v9's default retention policy.
1222 snapshot_cap_explicit: bool,
1223 /// FMT-9c: this flush wrote big runs straight to their home pages,
1224 /// which must be fsynced before the commit.
1225 direct_writes: bool,
1226 /// FMT-9b: keep the column resident while it fits this many bytes.
1227 vector_cache_budget: usize,
1228}
1229
1230/// Composite key for the lazy `key_index` on `Mnemo`. A slot is identified
1231/// by `(scope_owner, key)`:
1232/// * `Scope::Private` writes carry the owning agent → `Some(agent_id)`.
1233/// * `Scope::Shared` writes are per-file → `None`.
1234///
1235/// A private write never supersedes a shared one, and the reverse is also
1236/// true. Scope lives on `Memory.scope`, agent on `Memory.agent_id`.
1237#[derive(Debug, Clone, PartialEq, Eq, Hash)]
1238pub(crate) struct KeySlot {
1239 /// `Some(agent_id)` for `Scope::Private`, `None` for `Scope::Shared`.
1240 pub(crate) scope_owner: Option<String>,
1241 /// The caller-supplied key string, exactly as given (no case-fold).
1242 pub(crate) key: String,
1243}
1244
1245/// True if a memory's metadata marks it as the canonical onboarding manifest
1246/// (`metadata.topic == "manifest"`). Used by [`Mnemo::about`] to hoist the
1247/// manifest to the top of the briefing regardless of importance ordering.
1248fn is_manifest(m: &Memory) -> bool {
1249 m.metadata
1250 .get("topic")
1251 .and_then(|v| v.as_str())
1252 .map(|s| s.eq_ignore_ascii_case("manifest"))
1253 .unwrap_or(false)
1254}
1255
1256/// PERF-1 step 4: ULID → catalog row, as an open-addressing table of `u32`
1257/// row numbers: 4 bytes a slot, about 5 bytes an entry, against about 43
1258/// for the `HashMap<u128, usize>` it replaces. A probe compares against the
1259/// catalog's own id at that row, so the ids aren't stored twice. Rows are
1260/// never removed (a delete only tombstones the catalog entry), so there's
1261/// no delete path.
1262#[derive(Debug, Clone, Default)]
1263struct IdIndex {
1264 slots: Vec<u32>,
1265 len: usize,
1266}
1267
1268const NO_ROW: u32 = u32::MAX;
1269
1270impl IdIndex {
1271 fn build(catalog: &[CatalogEntry]) -> IdIndex {
1272 let mut t = IdIndex {
1273 slots: vec![NO_ROW; crate::index::table_capacity_for(catalog.len())],
1274 len: 0,
1275 };
1276 for (row, e) in catalog.iter().enumerate() {
1277 t.insert(e.id, row, catalog);
1278 }
1279 t
1280 }
1281
1282 /// The catalog row holding `id`, if any.
1283 fn get(&self, id: u128, catalog: &[CatalogEntry]) -> Option<usize> {
1284 if self.slots.is_empty() {
1285 return None;
1286 }
1287 let cap = self.slots.len();
1288 let mut i = crate::index::slot_home(id, cap);
1289 loop {
1290 let row = self.slots[i];
1291 if row == NO_ROW {
1292 return None;
1293 }
1294 if catalog[row as usize].id == id {
1295 return Some(row as usize);
1296 }
1297 i = crate::index::next_slot(i, cap);
1298 }
1299 }
1300
1301 /// Map `id` to `row`. `catalog` must hold every row already in the
1302 /// table (the new row may not be pushed yet).
1303 fn insert(&mut self, id: u128, row: usize, catalog: &[CatalogEntry]) {
1304 if (self.len + 1) * 4 > self.slots.len() * 3 {
1305 let cap = (self.slots.len() * 2).max(crate::index::table_capacity_for(self.len + 1));
1306 let old = std::mem::replace(&mut self.slots, vec![NO_ROW; cap]);
1307 self.len = 0;
1308 for r in old.into_iter().filter(|&r| r != NO_ROW) {
1309 self.insert(catalog[r as usize].id, r as usize, catalog);
1310 }
1311 }
1312 let cap = self.slots.len();
1313 let mut i = crate::index::slot_home(id, cap);
1314 loop {
1315 let r = self.slots[i];
1316 if r == NO_ROW {
1317 self.slots[i] = row as u32;
1318 self.len += 1;
1319 return;
1320 }
1321 if catalog[r as usize].id == id {
1322 self.slots[i] = row as u32;
1323 return;
1324 }
1325 i = crate::index::next_slot(i, cap);
1326 }
1327 }
1328
1329 #[cfg(feature = "dev-hooks")]
1330 fn heap_bytes(&self) -> usize {
1331 self.slots.capacity() * std::mem::size_of::<u32>()
1332 }
1333}
1334
1335/// Build the ULID → catalog-row lookup.
1336fn build_id_index(catalog: &[CatalogEntry]) -> IdIndex {
1337 IdIndex::build(catalog)
1338}
1339
1340// ---------------------------------------------------------------------------
1341// Phase 12.7a helpers used by `ingest_thread`.
1342// ---------------------------------------------------------------------------
1343
1344/// PERF-1 / FMT-9: was this catalog row written after the persisted index
1345/// run? v8 allocates pages monotonically, so a record past the run's first
1346/// page is newer. v9 reuses pages, so it compares generations instead.
1347fn written_after_index(e: &CatalogEntry, v9: bool, index_start: u64, index_gen: u64) -> bool {
1348 if v9 {
1349 // `start_page` 0: packed at the next flush, so certainly newer.
1350 e.start_page == 0 || e.gen > index_gen
1351 } else {
1352 e.start_page > index_start
1353 }
1354}
1355
1356/// Whether a record belongs in the ANN index: facts only. Thread-archive
1357/// turns, essence items and thread links are read by kind and scope, never
1358/// by similarity, so they stay out (R-004.2).
1359/// v0.16.1: a thread turn whose content is only transcript markers
1360/// ([`MNEMO_META_MARKERS_ONLY`]), by its flag or, for one saved before the
1361/// flag, by its content.
1362pub(super) fn is_markers_only_turn(m: &Memory) -> bool {
1363 meta_str(m, MNEMO_META_KIND) == Some(MNEMO_META_KIND_TURN)
1364 && (m.metadata.get(MNEMO_META_MARKERS_ONLY) == Some(&Value::Bool(true))
1365 || crate::transcript::is_marker_only(&m.content))
1366}
1367
1368#[cfg(test)]
1369#[test]
1370fn a_marker_only_turn_saved_before_the_flag_is_recognised_by_its_content() {
1371 let turn = |content: &str| {
1372 let mut m = Memory::new(content, MemoryType::Episodic, vec![1.0]);
1373 m.metadata.insert(
1374 MNEMO_META_KIND.into(),
1375 Value::String(MNEMO_META_KIND_TURN.into()),
1376 );
1377 m
1378 };
1379 assert!(is_markers_only_turn(&turn("[thinking]")));
1380 assert!(is_markers_only_turn(&turn(
1381 "[tool_use name=mcp__mnemo__remember]"
1382 )));
1383 assert!(!is_markers_only_turn(&turn(
1384 "staging-7\n[tool_use name=Bash]"
1385 )));
1386 // Only turns: a fact that reads like a marker is a fact.
1387 assert!(!is_markers_only_turn(&Memory::new(
1388 "[thinking]",
1389 MemoryType::Semantic,
1390 vec![1.0]
1391 )));
1392}
1393
1394/// 13a: a caller's `_mnemo.*` keys: [`crate::memory::CALLER_KEYS`] are
1395/// allowed and checked; any other is refused.
1396pub(super) fn check_caller_reserved(metadata: &Map<String, Value>) -> Result<()> {
1397 for (k, v) in metadata {
1398 if !k.starts_with(crate::memory::MNEMO_META_NS) {
1399 continue;
1400 }
1401 if !is_caller_key(k) {
1402 return Err(crud::reserved_key_error(k));
1403 }
1404 crate::memory::check_caller_key(k, v).map_err(MnemoError::Invalid)?;
1405 }
1406 Ok(())
1407}
1408
1409fn is_ann_kind(m: &Memory) -> bool {
1410 !matches!(
1411 m.metadata.get(MNEMO_META_KIND).and_then(|v| v.as_str()),
1412 Some(MNEMO_META_KIND_TURN)
1413 | Some(MNEMO_META_KIND_ESSENCE)
1414 | Some(MNEMO_META_KIND_THREAD_LINK)
1415 ) && !is_coordination(m)
1416}