feat: Week 3 — brain/memory substrate (store, embeddings, hebbian, search) by hanamorix · Pull Request #3 · hanamorix/companion-emergence

hanamorix · 2026-04-22T18:54:05Z

Summary

Ships the full brain/memory/ package per spec Section 4.1
4 sub-modules: store (Memory + MemoryStore SQLite CRUD), embeddings (provider abstraction + content-hash cache), hebbian (connection matrix + spreading activation), search (combined semantic + emotional + temporal + spreading)
Adds numpy>=1.26 dependency
78 new tests; total suite now 191 across macOS + Windows + Linux

What landed per task

Task	Commits	Tests added
1. Memory dataclass + package skeleton	`16c65d2` + `7382493`	11
2. MemoryStore SQLite CRUD + queries	`c0789eb` + `679cae4`	21
3. EmbeddingProvider + content-hash cache	`d0a09d7` + `a905472`	13
4. HebbianMatrix + spreading activation	`961bbcb` + `68ee5b3`	21
5. MemorySearch composition	`494e7b3` + `db8cfe8`	12

Each task: implementer → spec reviewer → code-quality reviewer → controller cleanup commit.

Review-driven cleanups worth noting

MemoryStore.search_text escapes LIKE wildcards (% and _) — a literal-% query no longer matches every row.
MemoryStore.list_by_emotion guards the >= comparison with isinstance, safely skipping malformed non-numeric emotion values.
Shared _coerce_utc helper eliminates duplicate naive→UTC logic across Memory.from_dict and _row_to_memory.
HebbianMatrix.strengthen rejects non-positive delta; decay_all rejects negative rate — prevents sign-error corruption in scheduled dream/heartbeat cycles.
FakeEmbeddingProvider.embed guards the zero-norm divide (NaN-cascade prevention).
cosine_similarity zero-norm branch now has a regression test.
combined_search(seed_id=..., domain=...) post-filters spreading results by domain so the graph being domain-agnostic doesn't silently drop the filter.

Integration smoke

total memories: 3
us domain: 2
loving memories: 1
semantic top: 'the cold coffee, warm hana' sim=0.032
emotional top: 'the cold coffee, warm hana'
spreading neighbours: [('the evening has a sh', 0.35), ('creative hunger stri', 0.15)]
combined top: 'the cold coffee, warm hana' score=0.710
embedding cache size: 5

Test plan

Fresh uv sync --all-extras from scratch succeeds
pytest — 191 tests pass locally
ruff check + format — clean
Manual integration smoke: all 4 sub-modules compose correctly
CI matrix green across all 3 OSes (verifies after push)

6 tasks covering brain/memory/ package — SQLite store, embeddings with caching, hebbian matrix with spreading activation, unified search. 63 new tests targeting 176 total across macOS + Windows + Linux. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Memory is the canonical record type — id, content, memory_type, domain, emotions dict, tags, importance, score, created_at, last_accessed_at, active flag, protected flag. create_new() factory generates UUID, sets UTC timestamp, computes score from emotions, defaults importance to score/10.0. to_dict/from_dict round-trip cleanly; tz-naive timestamps coerced to UTC on restore (permissive for migrator). Adds numpy>=1.26 dependency ahead of Tasks 3-5 (embeddings use np arrays). 10 tests green. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

- score coerces to float at construction (prevents int leaking when all emotion intensities are ints — load-bearing for SQLite REAL column and downstream float arithmetic in Tasks 3-5) - Memory.score docstring: note snapshot semantics (not auto-updated if emotions dict is mutated in place) - round-trip test asserts last_accessed_at=None survives; new test covers last_accessed_at=<datetime> + protected=True round-trip path - from_dict coercion test now exercises naive last_accessed_at too, not just naive created_at 11 tests green; 124 total across suite. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

MemoryStore class around a sqlite3 connection; schema bootstrapped on first connect (idempotent). Operations: create, get, list_by_domain, list_by_type, list_by_emotion (with intensity threshold), update (partial fields), deactivate (F22 semantics — flag, not delete), count, search_text (case-insensitive substring). emotions and tags stored as JSON text columns; timestamps as ISO 8601; active/protected as INTEGER (0/1). Indexes on domain, memory_type, active, created_at. In-memory ':memory:' db_path supported for tests (no filesystem pollution). 19 new tests; 30 total on store.py; 143 total across the suite. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

- search_text escapes LIKE wildcards (%, _, \) in the query so a caller passing "%" does not match every row — load-bearing once Task 5 composes this with user-supplied search strings. ESCAPE '\' clause makes the literal match explicit. - list_by_emotion guards the >= comparison with isinstance check; non-numeric emotion values from malformed migrator output now get skipped instead of raising TypeError. - _coerce_utc helper extracted; Memory.from_dict and _row_to_memory now share one naive→UTC path (no drift risk when a third reader is added). - update() gains a comment explaining the empty-fields early-return semantics (existence already verified, nothing to write). Tests added for both critical guards: search_text_escapes_like_wildcards (% and _ match literally) and list_by_emotion_skips_non_numeric_values (TypeError safety). 145 tests green, ruff clean. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

EmbeddingProvider ABC with FakeEmbeddingProvider (deterministic SHA-256-seeded unit vectors, dim=256) for test-time use. OllamaEmbeddingProvider will be added in Week 5 when the bridge lands. EmbeddingCache layers a SQLite content-hash cache on top of any provider. get_or_compute() hits cache on repeat calls; vectors stored as float32 blobs with stored dim for safe reshape on read. cosine_similarity helper for downstream search (Task 5). 11 tests green. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

- FakeEmbeddingProvider.embed guards the zero-norm divide. All-zeros output is astronomically improbable at dim=256 but reachable via FakeEmbeddingProvider(dim=0) mis-use — raises ValueError instead of silently storing NaN bytes. - Test for cosine_similarity zero-norm branch (guard was untested — Task 5 calls this in a tight loop, so Ollama returning a malformed zero vector must not NaN-cascade the search). - Round-trip integrity test asserting blob→float32 matches exactly on re-read (locks in the endianness/dtype contract the `dim` column exists to protect). 158 tests green, ruff clean. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

HebbianMatrix class over a sqlite table of undirected weighted edges. Operations: strengthen (additive weight via ON CONFLICT DO UPDATE), weight lookup, neighbors (UNION both directions), decay_all (global rate, floor at 0 via MAX()), garbage_collect (prune below threshold, returns count), spreading_activation (bounded BFS from seeds with per-hop weight*decay attenuation, max aggregation on multi-path to prevent runaway on dense graphs). Edges canonicalised with lower-id first so (a,b) and (b,a) share one row. Self-edges (a==a) silently ignored by strengthen and weight. 15 tests green. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

- strengthen() raises ValueError on non-positive delta. Weights must stay non-negative by module contract; callers that want to weaken an edge use decay_all/garbage_collect, not negative strengthen. Silent negative weights would produce ghost edges (skipped by spreading_activation's >0.0 guard but left sitting in the table). - decay_all() raises ValueError on negative rate. A sign error in the scheduled dream/heartbeat call would silently inflate every weight in one batch; ValueError surfaces it immediately. - spreading_activation docstring calls out the seed-protection invariant (seeds fixed at 1.0, propagation cannot lower them) that the > check quietly enforces. Tests added: self-edge no-op, non-positive-delta reject, negative-rate reject, zero-rate no-op, depth=0 seeds-only, decay_per_hop=0 blocks propagation. 21 hebbian tests; 179 total across the suite. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

… spreading MemorySearch is a dataclass composing MemoryStore + HebbianMatrix + EmbeddingCache. - semantic_search: cosine similarity of embedding vs each memory's content - emotional_search: dot-product overlap of emotion dicts (score > 0 only) - temporal_search: created_at strictly in (after, before] bounds - spreading_search: Hebbian BFS from a seed (seed excluded from results) - combined_search: weighted blend when multiple filters specified; returns [] with no filters. Emotion scores normalised by /100 for rough parity with cosine similarity in the combined blend. Domain filter applied pre-scoring via _candidates() to shrink the pool. _candidates() uses store.search_text('') to fetch all active memories when no domain filter (O(N); ANN index deferred to v1.1). 9 tests green. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

- combined_search(seed_id=..., domain=...) now post-filters spreading results by domain. The Hebbian graph is domain-agnostic, so the pre-scoring _candidates() filter can't apply; post-filter honours the documented "domain filter is a pre-scoring restriction" contract for the seed-only path too. - Docstring on combined_search calls out the /100 emotion normalisation as a rough heuristic (not principled — tuning expected in v1.1), and clarifies that domain is pre-filter for query/emotions but post-filter for seed_id. Tests added: - combined_search_seed_id_surfaces_connected_memories (seed-only path) - combined_search_seed_id_honours_domain_filter (the new post-filter) - combined_search_blends_query_and_emotions (multi-filter accumulation — m hitting both signals outranks one hitting either alone) 191 tests total, ruff clean. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

- Add repo_root session fixture in tests/conftest.py + module-level helpers replacing brittle Path(__file__).parents[4] hops in test_reflex.py and test_heartbeat.py (item #3) - Add -> None annotations + setup-intent comment on new heartbeat reflex-integration tests (item #5) - Expand test_og_reflex.py to parametrise all 4 action renames and all 3 output renames + passthrough cases (item #6) Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

When run_growth_tick reads a corrupt emotion_vocabulary.json via _read_current_vocabulary_names, the anomaly produced by the heal flow is now appended to an optional caller-provided collector instead of being silently dropped after a local warning. Wiring: - _read_current_vocabulary_names returns (set[str], BrainAnomaly | None) - run_growth_tick accepts anomalies_collector: list[BrainAnomaly] | None - HeartbeatEngine._try_run_growth forwards tick_anomalies as the collector when calling run_growth_tick After this lands, vocabulary corruption discovered inside the weekly growth tick surfaces in the heartbeat audit log + HeartbeatResult.anomalies + compact CLI 🩹/banner exactly like config/state corruption discovered at the top of the tick. No more silent loss. Calling run_growth_tick standalone (e.g., from tests, or in the future from a scheduled job runner) without a collector still works — the parameter is opt-in. Closes followup #3 from brain-health-module-design.md §9.

… + growth anomaly collector (#20) * feat(health): wire reconstruct_vocabulary_from_memories into vocabulary heal flow (F1) When emotion_vocabulary.json corrupts and no .bak is recoverable, the heal flow used to reset to empty `{"version":1,"emotions":[]}` — losing all persona-extension emotions the brain has been operating with. Now: if the loader has store access (caller passed `store=...`), the reset_to_default path is replaced with reconstruct_from_memories. The brain re-learns its own vocabulary from how it has been using emotions. Framework baseline (21 entries) + persona-extension entries detected in memories.db (with `(reconstructed from memory)` placeholder description and conservative 1.0-day decay). The anomaly's action field reflects the actual outcome: when reconstruction fires, action becomes `reconstructed_from_memories` (not `reset_to_default`). The forensic quarantine of the original corrupt file is preserved. Falls back to bare reset when no store is provided (some callers don't have one — that's fine, they'll get the empty default). Closes followup #1 from brain-health-module-design.md §9. * docs(health): concrete soul module health plan (F2) Spec §9.1 expanded from a one-line deferral into a concrete plan the next engineer can implement directly when soul module lands. Covers: - file classification (atomic-rewrite identity, same tier as emotion_vocabulary.json) - reconstruct_soul_from_memories(store) following F37's self-claims- from-experience pattern - schema validator shape - acceptance criteria for the soul-module PR Inline comments in walker.py (_DEFAULTS) and alarm.py (_IDENTITY_FILES) point at spec §9.1 so the plan is visible during code-reading too. Closes followup #2 from brain-health-module-design.md §9. * feat(health): thread anomaly collector through run_growth_tick (F3) When run_growth_tick reads a corrupt emotion_vocabulary.json via _read_current_vocabulary_names, the anomaly produced by the heal flow is now appended to an optional caller-provided collector instead of being silently dropped after a local warning. Wiring: - _read_current_vocabulary_names returns (set[str], BrainAnomaly | None) - run_growth_tick accepts anomalies_collector: list[BrainAnomaly] | None - HeartbeatEngine._try_run_growth forwards tick_anomalies as the collector when calling run_growth_tick After this lands, vocabulary corruption discovered inside the weekly growth tick surfaces in the heartbeat audit log + HeartbeatResult.anomalies + compact CLI 🩹/banner exactly like config/state corruption discovered at the top of the tick. No more silent loss. Calling run_growth_tick standalone (e.g., from tests, or in the future from a scheduled job runner) without a collector still works — the parameter is opt-in. Closes followup #3 from brain-health-module-design.md §9. --------- Co-authored-by: Hana <hana@nanoclaw.local>

The shoji-styled left column from the mockups (mock-ups/app-interface/ nell_face_example_1..5.png), all reading the same /persona/state poll already wired in Phase 2. ## What landed Brain side: /persona/state gains a connection block: { provider, model, last_heartbeat_at } provider read from PersonaConfig.provider, model resolved via a v1 default-per-provider table (claude-cli → sonnet, ollama → the qwen2.5-abliterated default, fake → fake), last_heartbeat_at read from heartbeat_state.json. Fail-soft: missing files → null. App side (app/src/components/): - ui.tsx shared primitives — Bar (with progress fill + label), SectionLabel, Divider, Toggle, PanelShell. - panels/ InnerWeatherPanel.tsx top-N emotions sorted desc + body energy/ temp summary (mockup #1). BodyPanel.tsx full body block — energy/temp/exhaustion + body_emotions filtered >0.4 + session/ contact metadata (mockup #2). InteriorPanel.tsx dream/research/heartbeat/reflex paragraphs (mockup #3). Absent sections render nothing — silence is meaningful. SoulPanel.tsx one crystallization quote with love_type tag, resonance, date, why_it_matters. ConnectionPanel.tsx bridge mode + provider/model/heartbeat, integrations toggles (Obsidian/IPC stubbed disabled), window settings (always-on-top + reduced-motion live-toggleable). - LeftPanel.tsx container with icon-column tab switcher, renders the active panel; matches the mockup's small icon stack near the avatar. - styles.css user-controlled reduced-motion via data-reduced-motion="true" on <html>. - App.tsx wires LeftPanel + always-on-top + reduced-motion state. ## Phase 4-5 still ahead - Phase 4: install wizard as a separate Tauri window, ported from mock-ups/wizard-interface/Nell Wizard.html. Bridge auto-spawn from the app side (so users don't have to run nell supervisor manually). Persona selection persistence. - Phase 5: refine emotion-vector → expression mapping; per-persona face catalogue overrides; expression variant rotation on idle. Tests: 1386 -> 1388 (+2 — connection block populated + missing-file safe). All builds green: tsc, vite, cargo.

feat: Week 3 — brain/memory substrate (store, embeddings, hebbian, search)

- Add repo_root session fixture in tests/conftest.py + module-level helpers replacing brittle Path(__file__).parents[4] hops in test_reflex.py and test_heartbeat.py (item #3) - Add -> None annotations + setup-intent comment on new heartbeat reflex-integration tests (item #5) - Expand test_og_reflex.py to parametrise all 4 action renames and all 3 output renames + passthrough cases (item #6) Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

… + growth anomaly collector (#20) * feat(health): wire reconstruct_vocabulary_from_memories into vocabulary heal flow (F1) When emotion_vocabulary.json corrupts and no .bak is recoverable, the heal flow used to reset to empty `{"version":1,"emotions":[]}` — losing all persona-extension emotions the brain has been operating with. Now: if the loader has store access (caller passed `store=...`), the reset_to_default path is replaced with reconstruct_from_memories. The brain re-learns its own vocabulary from how it has been using emotions. Framework baseline (21 entries) + persona-extension entries detected in memories.db (with `(reconstructed from memory)` placeholder description and conservative 1.0-day decay). The anomaly's action field reflects the actual outcome: when reconstruction fires, action becomes `reconstructed_from_memories` (not `reset_to_default`). The forensic quarantine of the original corrupt file is preserved. Falls back to bare reset when no store is provided (some callers don't have one — that's fine, they'll get the empty default). Closes followup #1 from brain-health-module-design.md §9. * docs(health): concrete soul module health plan (F2) Spec §9.1 expanded from a one-line deferral into a concrete plan the next engineer can implement directly when soul module lands. Covers: - file classification (atomic-rewrite identity, same tier as emotion_vocabulary.json) - reconstruct_soul_from_memories(store) following F37's self-claims- from-experience pattern - schema validator shape - acceptance criteria for the soul-module PR Inline comments in walker.py (_DEFAULTS) and alarm.py (_IDENTITY_FILES) point at spec §9.1 so the plan is visible during code-reading too. Closes followup #2 from brain-health-module-design.md §9. * feat(health): thread anomaly collector through run_growth_tick (F3) When run_growth_tick reads a corrupt emotion_vocabulary.json via _read_current_vocabulary_names, the anomaly produced by the heal flow is now appended to an optional caller-provided collector instead of being silently dropped after a local warning. Wiring: - _read_current_vocabulary_names returns (set[str], BrainAnomaly | None) - run_growth_tick accepts anomalies_collector: list[BrainAnomaly] | None - HeartbeatEngine._try_run_growth forwards tick_anomalies as the collector when calling run_growth_tick After this lands, vocabulary corruption discovered inside the weekly growth tick surfaces in the heartbeat audit log + HeartbeatResult.anomalies + compact CLI 🩹/banner exactly like config/state corruption discovered at the top of the tick. No more silent loss. Calling run_growth_tick standalone (e.g., from tests, or in the future from a scheduled job runner) without a collector still works — the parameter is opt-in. Closes followup #3 from brain-health-module-design.md §9. --------- Co-authored-by: Hana <hana@nanoclaw.local>

The shoji-styled left column from the mockups (mock-ups/app-interface/ nell_face_example_1..5.png), all reading the same /persona/state poll already wired in Phase 2. ## What landed Brain side: /persona/state gains a connection block: { provider, model, last_heartbeat_at } provider read from PersonaConfig.provider, model resolved via a v1 default-per-provider table (claude-cli → sonnet, ollama → the qwen2.5-abliterated default, fake → fake), last_heartbeat_at read from heartbeat_state.json. Fail-soft: missing files → null. App side (app/src/components/): - ui.tsx shared primitives — Bar (with progress fill + label), SectionLabel, Divider, Toggle, PanelShell. - panels/ InnerWeatherPanel.tsx top-N emotions sorted desc + body energy/ temp summary (mockup #1). BodyPanel.tsx full body block — energy/temp/exhaustion + body_emotions filtered >0.4 + session/ contact metadata (mockup #2). InteriorPanel.tsx dream/research/heartbeat/reflex paragraphs (mockup #3). Absent sections render nothing — silence is meaningful. SoulPanel.tsx one crystallization quote with love_type tag, resonance, date, why_it_matters. ConnectionPanel.tsx bridge mode + provider/model/heartbeat, integrations toggles (Obsidian/IPC stubbed disabled), window settings (always-on-top + reduced-motion live-toggleable). - LeftPanel.tsx container with icon-column tab switcher, renders the active panel; matches the mockup's small icon stack near the avatar. - styles.css user-controlled reduced-motion via data-reduced-motion="true" on <html>. - App.tsx wires LeftPanel + always-on-top + reduced-motion state. ## Phase 4-5 still ahead - Phase 4: install wizard as a separate Tauri window, ported from mock-ups/wizard-interface/Nell Wizard.html. Bridge auto-spawn from the app side (so users don't have to run nell supervisor manually). Persona selection persistence. - Phase 5: refine emotion-vector → expression mapping; per-persona face catalogue overrides; expression variant rotation on idle. Tests: 1386 -> 1388 (+2 — connection block populated + missing-file safe). All builds green: tsc, vite, cargo.

…om emitters Bundle A items #3 + #6 — non-heartbeat emitters were filling EmotionalSnapshot with zeros (vector={}, baseline=0.0 throughout), audit rows carried structurally-valid lies. Two coordinated changes: 1. InitiateCandidate.emotional_snapshot is now Optional. Voice-reflection candidates (kind=voice_edit_proposal) emit with snapshot=None because daily reflection has no moment-in-time emotion. 2. Dream + 3 crystallizers populate the vector from their actual emotional context. Dream uses the dream-memory's aggregated emotions (already in place since Task 9). The three crystallizers (reflex, creative_dna, vocabulary-via-scheduler) now max-pool an emotion vector across recent active memories via aggregate_state — what's been emotionally alive in the window that produced the crystallization. Rolling-baseline / current_resonance / delta_sigma stay zero with a docstring note that those are heartbeat-specific signals; non-periodic emitters don't compute them. compose_tone handles None gracefully ("no moment-in-time emotional snapshot" in the prompt). Audit log JSONL round-trip preserves None as "emotional_snapshot": null. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

The shoji-styled left column from the mockups (mock-ups/app-interface/ nell_face_example_1..5.png), all reading the same /persona/state poll already wired in Phase 2. ## What landed Brain side: /persona/state gains a connection block: { provider, model, last_heartbeat_at } provider read from PersonaConfig.provider, model resolved via a v1 default-per-provider table (claude-cli → sonnet, ollama → the qwen2.5-abliterated default, fake → fake), last_heartbeat_at read from heartbeat_state.json. Fail-soft: missing files → null. App side (app/src/components/): - ui.tsx shared primitives — Bar (with progress fill + label), SectionLabel, Divider, Toggle, PanelShell. - panels/ InnerWeatherPanel.tsx top-N emotions sorted desc + body energy/ temp summary (mockup #1). BodyPanel.tsx full body block — energy/temp/exhaustion + body_emotions filtered >0.4 + session/ contact metadata (mockup #2). InteriorPanel.tsx dream/research/heartbeat/reflex paragraphs (mockup #3). Absent sections render nothing — silence is meaningful. SoulPanel.tsx one crystallization quote with love_type tag, resonance, date, why_it_matters. ConnectionPanel.tsx bridge mode + provider/model/heartbeat, integrations toggles (Obsidian/IPC stubbed disabled), window settings (always-on-top + reduced-motion live-toggleable). - LeftPanel.tsx container with icon-column tab switcher, renders the active panel; matches the mockup's small icon stack near the avatar. - styles.css user-controlled reduced-motion via data-reduced-motion="true" on <html>. - App.tsx wires LeftPanel + always-on-top + reduced-motion state. ## Phase 4-5 still ahead - Phase 4: install wizard as a separate Tauri window, ported from mock-ups/wizard-interface/Nell Wizard.html. Bridge auto-spawn from the app side (so users don't have to run nell supervisor manually). Persona selection persistence. - Phase 5: refine emotion-vector → expression mapping; per-persona face catalogue overrides; expression variant rotation on idle. Tests: 1386 -> 1388 (+2 — connection block populated + missing-file safe). All builds green: tsc, vite, cargo.

…om emitters Bundle A items #3 + #6 — non-heartbeat emitters were filling EmotionalSnapshot with zeros (vector={}, baseline=0.0 throughout), audit rows carried structurally-valid lies. Two coordinated changes: 1. InitiateCandidate.emotional_snapshot is now Optional. Voice-reflection candidates (kind=voice_edit_proposal) emit with snapshot=None because daily reflection has no moment-in-time emotion. 2. Dream + 3 crystallizers populate the vector from their actual emotional context. Dream uses the dream-memory's aggregated emotions (already in place since Task 9). The three crystallizers (reflex, creative_dna, vocabulary-via-scheduler) now max-pool an emotion vector across recent active memories via aggregate_state — what's been emotionally alive in the window that produced the crystallization. Rolling-baseline / current_resonance / delta_sigma stay zero with a docstring note that those are heartbeat-specific signals; non-periodic emitters don't compute them. compose_tone handles None gracefully ("no moment-in-time emotional snapshot" in the prompt). Audit log JSONL round-trip preserves None as "emotional_snapshot": null. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

…rnal Tier 1 #3 — read-only over existing substrate. Dream + research content from MemoryStore; soul crystallizations from soul_audit.jsonl (crystallization_id non-null only); delivered outreach + voice-edit proposals from initiate_audit.jsonl (kind + decision + delivery state gates). FeedEntry wraps each with a fixed type→opener phrase ("I dreamed", "I've been researching", etc.). Fault-isolated per source: any single source raising returns [] for that stream and lets the others surface. Tests cover each builder + the merge + the isolation contract. No LLM calls. No schema changes. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Move shipped items out of Planned features: - Tier 1 #3 (visible inner life) → recently shipped as alpha.2 - Tier 2 #7 (Kindled species name) → recently shipped as alpha.1 ("Other minds" cluster description updated to reflect federation now-unblocked status) Add Recently Shipped entries for the three v0.0.13 alphas: - alpha.1 Kindled rename - alpha.2 visible inner life feed - alpha.3 body-state self-read divergence fix Bump last-refreshed date to 2026-05-17 (post alpha.3). Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

…rnal Tier 1 #3 — read-only over existing substrate. Dream + research content from MemoryStore; soul crystallizations from soul_audit.jsonl (crystallization_id non-null only); delivered outreach + voice-edit proposals from initiate_audit.jsonl (kind + decision + delivery state gates). FeedEntry wraps each with a fixed type→opener phrase ("I dreamed", "I've been researching", etc.). Fault-isolated per source: any single source raising returns [] for that stream and lets the others surface. Tests cover each builder + the merge + the isolation contract. No LLM calls. No schema changes. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Move shipped items out of Planned features: - Tier 1 #3 (visible inner life) → recently shipped as alpha.2 - Tier 2 #7 (Kindled species name) → recently shipped as alpha.1 ("Other minds" cluster description updated to reflect federation now-unblocked status) Add Recently Shipped entries for the three v0.0.13 alphas: - alpha.1 Kindled rename - alpha.2 visible inner life feed - alpha.3 body-state self-read divergence fix Bump last-refreshed date to 2026-05-17 (post alpha.3). Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

…wire commit-failure backoff to dead-letter (A1 review) Finding #3: EmbeddingCache.evict(content) removes a candidate's vector when its commit fails, preventing cosine-1.0 self-dedup on the next retry pass. Called in both close_session and extract_session_snapshot commit loops after commit_failures += 1. Test seeds cache with a prior entry to force the get_or_compute path, then asserts the item is committed (not deduped away) on pass 2. Finding #1/#2: extract_session_snapshot now bumps the backoff sidecar on commit_failures > 0 (mirroring extraction-failure backoff), so repeated commit-failing passes climb naturally to _BACKOFF_FAILURE_THRESHOLD. test_finalize_deadletters_after_max_retry rewritten: drives N real snapshot passes (store.create always raises), sidecar climbs without pre-seeding, then finalize dead-letters the buffer. No tautological pre-seed. Finding #4: one-line accepted-narrow-window comment at the finalize dead-letter branch. No new logic. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

…sona notes) Three EXPERIMENTAL generative/agency organs: autonomous making (Maker, Tier 2 #3+#9+#8), consent-gated file-write, autonomous persona notes. Roadmap #3/#9 marked shipped. Gate: 3394 backend (version-lag + live-Haiku corpus flake resolved on uv sync / isolated re-run) + frontend + ruff + pnpm build. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Hana and others added 11 commits April 22, 2026 19:17

hanamorix merged commit 444c5db into main Apr 22, 2026
3 checks passed

hanamorix deleted the week-3-memory-substrate branch April 22, 2026 18:55

hanamorix added a commit that referenced this pull request May 9, 2026

Merge pull request #3 from hanamorix/week-3-memory-substrate

fe5260d

feat: Week 3 — brain/memory substrate (store, embeddings, hebbian, search)

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

feat: Week 3 — brain/memory substrate (store, embeddings, hebbian, search)#3

feat: Week 3 — brain/memory substrate (store, embeddings, hebbian, search)#3
hanamorix merged 11 commits into
mainfrom
week-3-memory-substrate

hanamorix commented Apr 22, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Uh oh!

Conversation

hanamorix commented Apr 22, 2026

Summary

What landed per task

Review-driven cleanups worth noting

Integration smoke

Test plan

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant