Skip to main content

mnemo_engine/store/
recall.rs

1//! Retrieval: memories(), about(), exact search, recall (scoring, filters, budgets, the ANN and BM25 candidate paths).
2
3use super::*;
4
5/// The thread a recall filter pins with `_mnemo.thread` equality, if any.
6fn filter_thread_pin(f: &Filter) -> Option<&str> {
7    f.terms.iter().find_map(|t| match t {
8        FilterTerm::Eq(k, Value::String(s)) if k == MNEMO_META_THREAD => Some(s.as_str()),
9        _ => None,
10    })
11}
12
13/// Whether a recall filter can only match records that aren't in the ANN
14/// (R-004.2): it pins `_mnemo.thread`, or restricts `_mnemo.kind` to turns,
15/// essence items or thread links. Such recalls scan exactly instead.
16fn filter_bypasses_ann(f: &Filter) -> bool {
17    let non_ann = |v: &Value| {
18        matches!(
19            v.as_str(),
20            Some(MNEMO_META_KIND_TURN)
21                | Some(MNEMO_META_KIND_ESSENCE)
22                | Some(MNEMO_META_KIND_THREAD_LINK)
23        )
24    };
25    filter_thread_pin(f).is_some()
26        || f.terms.iter().any(|t| match t {
27            FilterTerm::Eq(k, v) => k == MNEMO_META_KIND && non_ann(v),
28            FilterTerm::In(k, vs) => {
29                k == MNEMO_META_KIND && !vs.is_empty() && vs.iter().all(non_ann)
30            }
31            _ => false,
32        })
33}
34
35impl Mnemo {
36    /// Return every live memory (used by tooling and compaction).
37    pub fn memories(&mut self) -> Result<Vec<Memory>> {
38        // FMT-9b: one pass over the vector pages instead of one per record.
39        self.ensure_column()?;
40        let entries: Vec<CatalogEntry> = self
41            .catalog
42            .iter()
43            .filter(|e| !e.deleted)
44            .cloned()
45            .collect();
46        let mut out = Vec::with_capacity(entries.len());
47        for e in &entries {
48            out.push(self.read_memory(e)?);
49        }
50        Ok(out)
51    }
52
53    /// 12.2 (`mnemo://recent`): the `n` most recently written live memories,
54    /// newest first. Only facts: no thread turns, essence items, thread
55    /// links or onboarding records, and nothing superseded or expired. A
56    /// keyed replacement is a new record, so it counts from when it was
57    /// written; an in-place [`Mnemo::update`] keeps its `created_at`. Reads
58    /// the catalog's rows and decrypts only the records it returns (on v9),
59    /// and doesn't count as an access.
60    pub fn recent(&mut self, n: usize) -> Result<Vec<Memory>> {
61        let now = crate::memory::now_secs();
62        let mut rows: Vec<(i64, CatalogEntry)> = Vec::new();
63        let mut unrowed: Vec<CatalogEntry> = Vec::new();
64        for e in self.catalog.iter().filter(|e| !e.deleted) {
65            match e.v9 {
66                Some(r) => {
67                    if r.kind != ROW_KIND_FACT
68                        || r.flags & ROW_SUPERSEDED != 0
69                        || (r.flags & ROW_HAS_TTL != 0 && now >= r.expires_at)
70                        || self.superseded.as_ref().is_some_and(|s| s.contains(&e.id))
71                    {
72                        continue;
73                    }
74                    rows.push((r.created_at, e.clone()));
75                }
76                // A v8 file (read-only) has no rows: its records are read.
77                None => unrowed.push(e.clone()),
78            }
79        }
80        rows.sort_by(|a, b| b.0.cmp(&a.0).then_with(|| b.1.id.cmp(&a.1.id)));
81        let keep = |m: &Memory| {
82            !m.metadata.contains_key(MNEMO_META_SUPERSEDED_BY)
83                && !m.metadata.contains_key(MNEMO_META_KIND)
84                && !is_onboarding(m)
85                && !m.is_expired(now)
86        };
87        let mut out = Vec::new();
88        for (_, e) in rows {
89            if out.len() >= n {
90                break;
91            }
92            let m = self.read_memory(&e)?;
93            if keep(&m) {
94                out.push(m);
95            }
96        }
97        for e in unrowed {
98            let m = self.read_memory(&e)?;
99            if keep(&m) {
100                out.push(m);
101            }
102        }
103        out.sort_by(|a, b| {
104            b.created_at
105                .cmp(&a.created_at)
106                .then_with(|| b.id.cmp(&a.id))
107        });
108        out.truncate(n);
109        Ok(out)
110    }
111
112    /// Return the database's **self-describing onboarding memories** — the
113    /// ones tagged `metadata.area = "onboarding"`. This is the engine-level
114    /// surface for the *single-file philosophy*: an agent who receives a
115    /// `.mnemo` file (and its passphrase) can call this to learn what the
116    /// file is, which embedder it expects, the recommended agent id, and any
117    /// other conventions the file's author chose to record — all without
118    /// needing any external documentation.
119    ///
120    /// Ordering: the canonical manifest (tag `metadata.topic = "manifest"`)
121    /// always comes first; everything else follows in `importance` descending,
122    /// then `created_at` ascending for deterministic results.
123    pub fn about(&mut self) -> Result<Vec<Memory>> {
124        // Phase 12.1: drop records that have been superseded by a newer
125        // keyed write. They still live in snapshot history and
126        // `history(key)` — just not here.
127        // R-004.2: read only the onboarding records, from the handle caches.
128        self.ensure_key_caches()?;
129        let ids: Vec<u128> = self
130            .layers
131            .as_ref()
132            .map(|l| l.onboarding.iter().copied().collect())
133            .unwrap_or_default();
134        let mut out: Vec<Memory> = Vec::new();
135        for e in self.live_entries(ids) {
136            let m = self.read_memory(&e)?;
137            if !m.metadata.contains_key(MNEMO_META_SUPERSEDED_BY) && is_onboarding(&m) {
138                out.push(m);
139            }
140        }
141        out.sort_by(|a, b| {
142            let a_manifest = is_manifest(a);
143            let b_manifest = is_manifest(b);
144            // Manifest topic wins first; then importance desc; then created_at asc.
145            b_manifest
146                .cmp(&a_manifest)
147                .then_with(|| b.importance.total_cmp(&a.importance))
148                .then_with(|| a.created_at.cmp(&b.created_at))
149        });
150        Ok(out)
151    }
152
153    /// The project essence for the `about` briefing (Phase 12.7d): the
154    /// [`EssenceScope::Project`] render at the default budget
155    /// ([`ESSENCE_DEFAULT_RENDER_BYTES`]), or `None` when the file has no
156    /// open project-scope item or timeline entry. The CLI and MCP `about`
157    /// append it after the manifest and onboarding memories; [`Mnemo::about`]
158    /// itself keeps returning records.
159    pub fn about_essence(&mut self) -> Result<Option<String>> {
160        let entries = self.essence_entries()?;
161        let mut any = false;
162        for e in &entries {
163            let m = self.read_memory(e)?;
164            let is_project_essence = m.metadata.get(MNEMO_META_KIND).and_then(|v| v.as_str())
165                == Some(MNEMO_META_KIND_ESSENCE)
166                && !m.metadata.contains_key(MNEMO_META_THREAD);
167            if is_project_essence && status_of(&m) != Some(essence_status::RESOLVED) {
168                any = true;
169                break;
170            }
171        }
172        if !any {
173            return Ok(None);
174        }
175        self.essence(EssenceScope::Project, ESSENCE_DEFAULT_RENDER_BYTES)
176            .map(Some)
177    }
178
179    // --- retrieval --------------------------------------------------------
180
181    /// Phase-1 brute-force search: rank live memories by raw similarity only.
182    /// Read-only — does not update access statistics.
183    pub fn search(
184        &mut self,
185        query: &[f32],
186        top_k: usize,
187        metric: Metric,
188    ) -> Result<Vec<(Memory, f32)>> {
189        if query.len() != self.dimensions {
190            return Err(MnemoError::DimensionMismatch {
191                expected: self.dimensions,
192                got: query.len(),
193            });
194        }
195        let now = memory::now_secs();
196        let entries: Vec<CatalogEntry> = self
197            .catalog
198            .iter()
199            .filter(|e| !e.deleted)
200            .cloned()
201            .collect();
202
203        let mut scored: Vec<(Memory, f32)> = Vec::new();
204        for e in &entries {
205            let m = self.read_memory(e)?;
206            if m.is_expired(now) {
207                continue;
208            }
209            let sim = memory::similarity(metric, query, &m.vector);
210            scored.push((m, sim));
211        }
212        scored.sort_by(|a, b| b.1.total_cmp(&a.1));
213        scored.truncate(top_k);
214        Ok(scored)
215    }
216
217    /// Phase-5 multi-signal recall: rank by
218    /// `α·similarity + β·recency + γ·importance + δ·ln(freq)`, with type and
219    /// agent filtering. Updates `accessed_at` / `access_count` on the
220    /// returned memories (persisted on the next `flush`).
221    ///
222    /// When an ANN index has been built ([`Mnemo::build_index`]), recall runs
223    /// the tiered IVF→PQ→rerank pipeline: it scores only the similarity-nearest
224    /// `n_rerank` candidates rather than every memory. Without an index it
225    /// scores every live memory exactly.
226    /// One pass of filter + score for [`Mnemo::recall`]. Factored out of
227    /// `recall` so the outer path can retry with a wider shortlist when
228    /// the type/agent/supersession/TTL filters starve the ANN-narrowed
229    /// result below `top_k` (Phase 12.1 short-results-on-indexed-path
230    /// fix).
231    ///
232    /// * `force_exact = false` → use the ANN index when one exists; fall
233    ///   back to a full scan only if none is built.
234    /// * `force_exact = true` → skip the ANN index and scan every live
235    ///   non-superseded record. Used by the outer path as the last-resort
236    ///   attempt when widening the shortlist would already touch half or
237    ///   more of the live non-superseded set.
238    ///
239    /// Returns an unsorted, un-truncated `Vec<RecallResult>`; the outer
240    /// `recall` sorts and truncates once.
241    pub(super) fn score_once(
242        &mut self,
243        req: &RecallRequest,
244        now: i64,
245        force_exact: bool,
246    ) -> Result<Vec<RecallResult>> {
247        // Candidate set: ANN-narrowed when an index exists and the
248        // caller hasn't forced an exact scan, else everything live.
249        // R-004.2: a filter pinning `_mnemo.thread` scans just that
250        // thread's records (from the handle caches); one that can only
251        // match non-fact kinds skips the ANN, which doesn't hold them.
252        // FMT-9d: keyword matches, normalised by the best one.
253        self.kw = None;
254        // 12.3 step 4: a recall by text uses its text when no keywords
255        // were given (hybrid by default); empty keywords opt out.
256        if let Some(text) = req.keywords_for(self.is_v9()).map(str::to_string) {
257            let scores = self.bm25_scores(&text, req.bm25_k1, req.bm25_b)?;
258            let max = scores.values().copied().fold(0.0f32, f32::max);
259            if max > 0.0 {
260                self.kw = Some(
261                    scores
262                        .into_iter()
263                        .map(|(row, s)| (self.catalog[row].id, s / max))
264                        .collect(),
265                );
266            } else {
267                self.kw = Some(std::collections::HashMap::new());
268            }
269        }
270        if let Some(thread) = filter_thread_pin(&req.filter) {
271            let thread = thread.to_string();
272            let entries = self.thread_entries(&thread)?;
273            return self.score_entries(req, now, &entries);
274        }
275        // Keyword-only: the candidates are the matches.
276        if req.query.is_empty() {
277            let ids: Vec<u128> = self
278                .kw
279                .as_ref()
280                .map_or(Vec::new(), |k| k.keys().copied().collect());
281            let entries: Vec<CatalogEntry> = ids
282                .into_iter()
283                .filter_map(|id| self.index.get(id, &self.catalog))
284                .map(|i| self.catalog[i].clone())
285                .filter(|e| !e.deleted)
286                .collect();
287            return self.score_entries(req, now, &entries);
288        }
289        let force_exact = force_exact || filter_bypasses_ann(&req.filter);
290        let mut entries: Vec<CatalogEntry> = match (&self.ann, force_exact) {
291            (Some(ann), false) => {
292                let n_rerank = self.rerank_for(ann, req);
293                let ids = ann.query(&req.query, req.n_probe, n_rerank);
294                ids.iter()
295                    .filter_map(|&id| self.index.get(id, &self.catalog))
296                    .map(|i| self.catalog[i].clone())
297                    .filter(|e| !e.deleted)
298                    .collect()
299            }
300            _ => self
301                .catalog
302                .iter()
303                .filter(|e| !e.deleted)
304                .cloned()
305                .collect(),
306        };
307        // FMT-9d: an ANN shortlist plus the keyword matches it missed.
308        if let Some(kw) = &self.kw {
309            if entries.len() < self.catalog.len() {
310                let have: std::collections::HashSet<u128> = entries.iter().map(|e| e.id).collect();
311                let extra: Vec<u128> = kw.keys().copied().filter(|id| !have.contains(id)).collect();
312                for id in extra {
313                    if let Some(i) = self.index.get(id, &self.catalog) {
314                        if !self.catalog[i].deleted {
315                            entries.push(self.catalog[i].clone());
316                        }
317                    }
318                }
319            }
320        }
321
322        self.score_entries(req, now, &entries)
323    }
324
325    /// FMT-9d: a candidate's relevance: its vector similarity plus ε times
326    /// its normalised keyword score (keyword-only requests: just the
327    /// keyword score). Returns `(relevance, reported similarity)`.
328    pub(super) fn relevance(&self, req: &RecallRequest, id: u128, sim: f32) -> (f32, f32) {
329        let kw = self
330            .kw
331            .as_ref()
332            .and_then(|k| k.get(&id).copied())
333            .unwrap_or(0.0);
334        if req.query.is_empty() {
335            (kw, kw)
336        } else {
337            (sim + req.effective_keyword_weight() * kw, sim)
338        }
339    }
340
341    /// Score one candidate set for [`Mnemo::score_once`].
342    pub(super) fn score_entries(
343        &mut self,
344        req: &RecallRequest,
345        now: i64,
346        entries: &[CatalogEntry],
347    ) -> Result<Vec<RecallResult>> {
348        // FMT-9b: score v9 candidates from their rows and the fp16 column,
349        // decrypting only what makes `top_k`. A metadata filter needs the
350        // record body, so it takes the general path below.
351        if self.is_v9() && req.filter.is_empty() {
352            return self.score_rows(req, now, entries);
353        }
354        let mut scored: Vec<RecallResult> = Vec::with_capacity(entries.len());
355        for e in entries {
356            // Phase 12.1: drop superseded records.
357            //
358            // Fast path: if the lazy `superseded` set has already been
359            // built by a prior keyed op on this handle, check it before
360            // decrypt — O(1), no AES work. On a fresh read-only handle
361            // it is typically `None`, so we fall through to the slow
362            // path (one `_mnemo.superseded_by` lookup per surviving
363            // record, O(k) over user metadata), and never trigger the
364            // full-file scan that `ensure_key_caches()` would do.
365            if self.superseded.as_ref().is_some_and(|s| s.contains(&e.id)) {
366                continue;
367            }
368            let m = self.read_memory(e)?;
369            if m.metadata.contains_key(MNEMO_META_SUPERSEDED_BY) {
370                continue;
371            }
372            if m.is_expired(now) {
373                continue;
374            }
375            if let Some(types) = &req.memory_types {
376                if !types.contains(&m.memory_type) {
377                    continue;
378                }
379            }
380            if let Some(agent) = &req.agent_id {
381                let visible = m.agent_id == *agent || m.scope == Scope::Shared;
382                if !visible {
383                    continue;
384                }
385            }
386            // Phase 12.6: pre-scoring filter (session_id + metadata
387            // conditions). Cheap when `req.filter.is_empty()`; the
388            // filter struct's `matches` is a short-circuit AND.
389            if !req.filter.is_empty() && !req.filter.matches(&m) {
390                continue;
391            }
392            // v0.16.1: a turn of nothing but markers (`[thinking]`, a
393            // tool call) is never a recall hit, whatever the filter: it
394            // would match any query weakly. Its content decides, so a turn
395            // saved before v0.16.1 is left out too.
396            if is_markers_only_turn(&m) {
397                continue;
398            }
399            // Phase 12.7: default-exclude the thread layers from fact
400            // recall: archive turns (12.7a), essence items (12.7b; read
401            // by scope through the render, not by similarity) and
402            // thread-link markers (12.7c; they carry an all-equal
403            // sentinel vector and would surface as junk hits). A caller
404            // that wants them opts in by placing any `_mnemo.*` term in
405            // the filter (equality, `$in`, `$exists`, or a range on
406            // `_mnemo.thread` / `_mnemo.kind` / `_mnemo.seq`).
407            if !req.filter.mentions_mnemo_meta()
408                && matches!(
409                    m.metadata.get(MNEMO_META_KIND).and_then(|v| v.as_str()),
410                    Some(MNEMO_META_KIND_TURN)
411                        | Some(MNEMO_META_KIND_ESSENCE)
412                        | Some(MNEMO_META_KIND_THREAD_LINK)
413                )
414            {
415                continue;
416            }
417            let sim = if req.query.is_empty() {
418                0.0
419            } else {
420                memory::similarity(req.metric, &req.query, &m.vector)
421            };
422            let (rel, sim) = self.relevance(req, e.id, sim);
423            let age_secs = (now - m.created_at) as f32;
424            let score = req
425                .weights
426                .score(rel, age_secs, m.importance, m.access_count);
427            scored.push(RecallResult {
428                memory: m,
429                score,
430                similarity: sim,
431            });
432        }
433        Ok(scored)
434    }
435
436    /// FMT-9b: [`Mnemo::score_entries`] for a v9 file with no metadata
437    /// filter. Every check the general path makes on the decrypted record
438    /// is made on the row instead (supersession, TTL, type, agent
439    /// visibility, the thread-layer kinds), the similarity comes from the
440    /// fp16 column, and the score from the row's `created_at` and
441    /// `importance`. Candidates are then decrypted best-first until `top_k`
442    /// survive; the agent check is repeated on the record, since the row
443    /// only has a hash of the agent id.
444    pub(super) fn score_rows(
445        &mut self,
446        req: &RecallRequest,
447        now: i64,
448        entries: &[CatalogEntry],
449    ) -> Result<Vec<RecallResult>> {
450        self.ensure_column()?;
451        let d = self.dimensions;
452        let q = &req.query;
453        let q_norm = q.iter().map(|x| x * x).sum::<f32>().sqrt();
454        let agent_h = req.agent_id.as_deref().map(|a| fnv1a32(&[a.as_bytes()]));
455        let types: Option<Vec<u8>> = req
456            .memory_types
457            .as_ref()
458            .map(|ts| ts.iter().map(|&t| memory_type_code(t)).collect());
459        let mut cand: Vec<(f32, f32, usize)> = Vec::with_capacity(entries.len());
460        for (pos, e) in entries.iter().enumerate() {
461            let Some(r) = e.v9 else { continue };
462            if r.flags & ROW_SUPERSEDED != 0 {
463                continue;
464            }
465            if self.superseded.as_ref().is_some_and(|s| s.contains(&e.id)) {
466                continue;
467            }
468            if r.flags & ROW_HAS_TTL != 0 && now >= r.expires_at {
469                continue;
470            }
471            if let Some(ts) = &types {
472                if !ts.contains(&r.memory_type) {
473                    continue;
474                }
475            }
476            if let Some(ah) = agent_h {
477                if r.flags & ROW_SHARED == 0 && r.agent_hash != ah {
478                    continue;
479                }
480            }
481            if matches!(
482                r.kind,
483                ROW_KIND_TURN | ROW_KIND_ESSENCE | ROW_KIND_THREAD_LINK
484            ) {
485                continue;
486            }
487            let Some(idx) = self.index.get(e.id, &self.catalog) else {
488                continue;
489            };
490            if !req.query.is_empty()
491                && self.col.resident
492                && idx < self.col.have.len()
493                && !self.col.have[idx]
494                && !self.col.hot.contains_key(&idx)
495            {
496                self.fill_column_page(idx)?;
497            }
498            let sim = if req.query.is_empty() {
499                0.0
500            } else if let Some((h, n)) = self.col.hot.get(&idx) {
501                veccol::similarity_f16(req.metric, q, q_norm, h, *n)
502            } else if self.col.resident && self.col.have.get(idx).copied().unwrap_or(false) {
503                veccol::similarity_f16(
504                    req.metric,
505                    q,
506                    q_norm,
507                    &self.col.data[idx * d..(idx + 1) * d],
508                    self.col.norms[idx],
509                )
510            } else {
511                let (h, n) = self.column_entry(e)?;
512                veccol::similarity_f16(req.metric, q, q_norm, &h, n)
513            };
514            let (rel, sim) = self.relevance(req, e.id, sim);
515            let score = req.weights.score(
516                rel,
517                (now - r.created_at) as f32,
518                r.importance,
519                e.access_count,
520            );
521            cand.push((score, sim, pos));
522        }
523        cand.sort_by(|a, b| b.0.total_cmp(&a.0));
524        let mut out: Vec<RecallResult> = Vec::with_capacity(req.top_k.min(cand.len()));
525        for (score, sim, pos) in cand {
526            if out.len() >= req.top_k {
527                break;
528            }
529            let m = self.read_memory(&entries[pos])?;
530            if let Some(agent) = &req.agent_id {
531                if !(m.agent_id == *agent || m.scope == Scope::Shared) {
532                    continue;
533                }
534            }
535            out.push(RecallResult {
536                memory: m,
537                score,
538                similarity: sim,
539            });
540        }
541        Ok(out)
542    }
543
544    /// 12.3 step 5: the keywords the registered embedder chooses for this
545    /// request ([`crate::Embedder::query_keywords`]): asked only for a recall
546    /// by text with no keywords given, on a v9 file (a v8 file has no
547    /// keyword index and ignores the text).
548    fn embedder_keywords(&self, req: &RecallRequest) -> Option<String> {
549        if req.keywords.is_some() || !self.is_v9() {
550            return None;
551        }
552        let text = req.text.as_deref().filter(|t| !t.trim().is_empty())?;
553        self.embedder.as_ref()?.query_keywords(text)
554    }
555
556    /// Phase-5 multi-signal recall under the v0.7.0 scoring-v2 formula
557    /// (see [`crate::ScoreWeights`]). Thin wrapper over
558    /// [`Mnemo::recall_budgeted`] that drops the `truncated` flag;
559    /// use the budgeted form if you need to know whether
560    /// [`RecallRequest::max_bytes`] cut anything.
561    pub fn recall(&mut self, req: &RecallRequest) -> Result<Vec<RecallResult>> {
562        Ok(self.recall_budgeted(req)?.hits)
563    }
564
565    /// Phase 12.8 budgeted recall: scoring v2 (see [`crate::ScoreWeights`])
566    /// with an optional UTF-8-content byte cap. Ranks by
567    /// `sim + max(sim, 0) · (β·recency + γ·(importance − 0.5) + δ·tanh(uses/5))`,
568    /// with type / agent / supersession / TTL filters and the 12.6
569    /// `Filter` applied before the scorer. Updates `accessed_at` /
570    /// `access_count` on the returned memories (persisted on the next
571    /// `flush`).
572    ///
573    /// When an ANN index has been built ([`Mnemo::build_index`]),
574    /// recall runs the tiered IVF→PQ→rerank pipeline. If the filters
575    /// (type / agent / supersession / TTL / 12.6) reduce the ANN
576    /// shortlist below `top_k`, Phase 12.1's short-results-on-indexed-
577    /// path fix kicks in: the query is retried with `n_probe` and
578    /// `n_rerank` doubled, and if that would already touch half or
579    /// more of the live non-superseded records, the ANN pass is
580    /// skipped for an exact scan. Without an index, recall scores
581    /// every live non-superseded memory.
582    ///
583    /// Phase 12.8: after `top_k` is applied, if
584    /// [`RecallRequest::max_bytes`] is set, hits are dropped in rank
585    /// order once the running sum of `content` UTF-8 bytes would
586    /// exceed the budget. The top-1 hit is always kept even if it
587    /// alone exceeds the budget. [`RecallResponse::truncated`] is
588    /// `true` iff at least one hit was dropped by the budget.
589    pub fn recall_budgeted(&mut self, req: &RecallRequest) -> Result<RecallResponse> {
590        // 12.3 step 5 (v0.15): the registered embedder may name the keywords
591        // for a recall by text (the built-in's identifier-aware recall).
592        let chosen;
593        let req = match self.embedder_keywords(req) {
594            Some(k) => {
595                chosen = req.clone().keywords(k);
596                &chosen
597            }
598            None => req,
599        };
600        // FMT-9d: keyword recall needs a v9 file's BM25 index; a
601        // keyword-only request has no vector.
602        if req
603            .keywords
604            .as_deref()
605            .is_some_and(|k| !k.trim().is_empty())
606            && !self.is_v9()
607        {
608            return Err(MnemoError::Invalid(
609                "keyword recall (keywords) needs a v9 file; run `mnemo compact` to upgrade".into(),
610            ));
611        }
612        let keyword_only = req.query.is_empty() && req.keywords.is_some();
613        if !keyword_only && req.query.len() != self.dimensions {
614            return Err(MnemoError::DimensionMismatch {
615                expected: self.dimensions,
616                got: req.query.len(),
617            });
618        }
619        // PERF-1 step 3: the ANN index loads on the first recall.
620        self.ensure_ann()?;
621        // Access-tracking bumps `accessed_at` / `access_count` in the
622        // catalog, which is a mutation. On a read-only handle the caller
623        // must pass `track_access(false)` (via CLI `--no-track`).
624        // 12.4: a shared read (read-only for the call, with a delta) keeps
625        // the bumps for the handle's next commit instead.
626        if self.read_only && req.track_access && self.access_delta.is_none() {
627            return Err(MnemoError::ReadOnly);
628        }
629        // Phase 12.1 **fix** (lane A v0.8.0 REVIEW): do NOT call
630        // `ensure_key_caches()` here. A fresh read-only handle that
631        // only ever recalls would otherwise pay a full-file decrypt
632        // on its first query to build the lazy superseded set. Instead
633        // `score_once` checks `_mnemo.superseded_by` per-record after
634        // decrypt (slow path) and consults the set only when it has
635        // already been built by a prior keyed op on this handle (fast
636        // path). `live_non_superseded` is not knowable without a
637        // decrypt pass either, so we approximate by `live_count`; the
638        // exact-scan gate uses this as a conservative upper bound,
639        // which errs toward widening (cheaper) rather than skipping
640        // straight to a full scan.
641        let live_count = self.catalog.iter().filter(|e| !e.deleted).count();
642        let live_non_superseded = live_count;
643        let now = memory::now_secs();
644
645        // Scoring v2 (shipped v0.7.0). Single-pass filter + score:
646        //   * `age_secs = now - m.created_at`, so a recall doesn't bump
647        //     its own results' recency on later queries.
648        //   * `ScoreWeights::score` applies the similarity-gated,
649        //     importance-centered, saturating-usage formula documented
650        //     on `ScoreWeights`.
651        //
652        // Phase 12.1 + 12.6 short-results-on-indexed-path fix: when the
653        // ANN shortlist is narrower than the live count, the filters
654        // (type / agent / supersession / TTL / 12.6) can reduce the
655        // result below `top_k`. If that happens, re-query the index
656        // with `n_probe` / `n_rerank` doubled from their *effective*
657        // base; if that would already cover ≥50% of the live set,
658        // skip the index and do an exact scan.
659        //
660        // Lane A v0.8.0 REVIEW fix: widen from
661        // `IvfPqIndex::effective_defaults()` when the caller didn't
662        // pin the knobs. The pre-fix code only doubled when `req.n_*`
663        // was already `Some(_)`, which skipped the widen step for the
664        // default (unpinned) recall path and jumped straight to an
665        // exact scan on moderately selective filters.
666        let mut scored = self.score_once(req, now, false)?;
667        let mut used_exact_scan = self.ann.is_none() || filter_bypasses_ann(&req.filter);
668        if !used_exact_scan && scored.len() < req.top_k && live_non_superseded > req.top_k {
669            if let Some(ann) = &self.ann {
670                let (default_probe, default_rerank) = ann.effective_defaults();
671                let base_probe = req.n_probe.unwrap_or(default_probe);
672                let base_rerank = self.rerank_for(ann, req).unwrap_or(default_rerank);
673                let widen_probe = base_probe.saturating_mul(2);
674                let widen_rerank = base_rerank.saturating_mul(2);
675                // Attempt 2: widen once, 2× from the effective base.
676                // Only worth another ANN pass if either knob actually
677                // grew (saturating_mul on usize::MAX would cap).
678                if widen_probe > base_probe || widen_rerank > base_rerank {
679                    let widened = RecallRequest {
680                        n_probe: Some(widen_probe),
681                        n_rerank: Some(widen_rerank),
682                        ..req.clone()
683                    };
684                    scored = self.score_once(&widened, now, false)?;
685                }
686                // Attempt 3: if still short after widening, pay the
687                // exact scan — it's the only path that guarantees
688                // `top_k` correct hits when the filter is more
689                // selective than `widen_rerank / live_count`. Before
690                // the fix this was gated on `widened_rerank * 2 >=
691                // live_non_superseded`, which silently returned fewer
692                // than `top_k` hits on selective filters that fell in
693                // the gap. Exact scan is bounded (one decrypt per
694                // live record); the previous widen is an optimization.
695                if scored.len() < req.top_k {
696                    scored = self.score_once(req, now, true)?;
697                    used_exact_scan = true;
698                }
699            }
700        }
701        scored.sort_by(|a, b| b.score.total_cmp(&a.score));
702        scored.truncate(req.top_k);
703
704        // Update access statistics for everything we surfaced — but only
705        // mutate the catalog entry, not the full record on disk. Pre-v5
706        // this loop called `self.put(r.memory.clone())`, which serialized
707        // and rewrote every result (vector included) to fresh pages,
708        // turning a top-K recall into K full record rewrites. Now the
709        // catalog is the source of truth for these two fields and the
710        // record body stays untouched until the next real edit.
711        if req.track_access && self.read_only {
712            if let Some(delta) = self.access_delta.as_mut() {
713                for r in &scored {
714                    delta.memory_hit(r.memory.id.0, now);
715                }
716            }
717        } else if req.track_access {
718            for r in &mut scored {
719                if let Some(idx) = self.index.get(r.memory.id.0, &self.catalog) {
720                    self.touch(idx);
721                    let entry = &mut self.catalog[idx];
722                    entry.accessed_at = now;
723                    entry.access_count = entry.access_count.saturating_add(1);
724                    // Propagate the just-updated values onto the returned
725                    // Memory so the caller sees the post-recall state.
726                    r.memory.accessed_at = entry.accessed_at;
727                    r.memory.access_count = entry.access_count;
728                }
729            }
730            // Mark the catalog dirty so the next flush persists these
731            // bumps. The records themselves are not dirty.
732            if !scored.is_empty() {
733                self.dirty_catalog = true;
734            }
735        }
736
737        // Phase 12.8: apply the `max_bytes` budget *after* tracking has
738        // seen the full top-k list. Access stats record what the scorer
739        // chose; the byte budget shouldn't suppress the signal.
740        let truncated = if let Some(budget) = req.max_bytes {
741            let orig_len = scored.len();
742            let mut running = 0usize;
743            let mut cut_at: Option<usize> = None;
744            for (i, r) in scored.iter().enumerate() {
745                let w = r.memory.content.len();
746                if i == 0 {
747                    // Always keep the top hit, even if it's over budget
748                    // on its own. Avoid "I asked for top_k results and
749                    // got zero" when a single memory exceeds the cap.
750                    running += w;
751                    continue;
752                }
753                if running.saturating_add(w) > budget {
754                    cut_at = Some(i);
755                    break;
756                }
757                running += w;
758            }
759            if let Some(idx) = cut_at {
760                scored.truncate(idx);
761                scored.len() < orig_len
762            } else {
763                false
764            }
765        } else {
766            false
767        };
768
769        self.kw = None;
770        Ok(RecallResponse {
771            hits: scored,
772            truncated,
773            used_exact_scan,
774        })
775    }
776
777    // --- approximate index ------------------------------------------------
778}