Changeset 0.18.0 (#79)

CHANGELOG.md

@@ -2,6 +2,90 @@
 
 All notable changes to Chat Switchboard.
 
+## [0.18.0] — 2026-02-28
+
+### Added
+- **Memory system.** Long-term memory across conversations with scope-aware
+  isolation. Three memory scopes: `user` (personal facts/preferences),
+  `persona` (shared across all users of a Persona), and `persona_user`
+  (per-user within a Persona context, e.g. tutoring progress per student).
+  Memories persist across channels and are injected into the system prompt
+  at completion time with scope priority (persona_user > persona > user).
+- **Memory tools.** Two new LLM-callable tools:
+  - `memory_save` — LLM explicitly stores a fact with key/value/confidence.
+    Scope-aware: saves to the active scope for the current Persona context.
+  - `memory_recall` — LLM queries stored facts relevant to current context.
+    Merges results from applicable scopes with semantic search via embeddings
+    and keyword fallback.
+- **Automatic memory extraction.** Background scanner finds conversations
+  with sufficient new activity, sends them to the utility model role for
+  fact extraction, and stores results as `pending_review` memories.
+  Configurable extraction prompt per Persona (e.g. "extract FAQ-worthy
+  Q&A pairs"). Scanner runs on a configurable interval with concurrency
+  control. Global kill switch via admin settings.
+- **Memory review pipeline.** Extracted memories start in `pending_review`
+  status. Admin panel shows pending review queue with bulk approve. Users
+  can approve/reject/edit their own pending memories from the Settings →
+  Memory tab.
+- **User memory management.** Settings → Memory tab shows active/pending
+  memory counts, filterable/searchable memory list with inline edit and
+  delete. Status filter (active, pending_review, archived). Approve All
+  button for batch operations on pending memories.
+- **Admin memory controls.** Admin panel → Memory section shows system-wide
+  pending review queue. Memory extraction toggle in admin Settings panel
+  (`memory_extraction_enabled`). Bulk approve endpoint for batch review.
+- **Hybrid semantic recall.** Memory recall supports both keyword search
+  and vector similarity (cosine distance). Postgres uses `<=>` operator
+  with pgvector; SQLite computes cosine similarity in Go with app-level
+  vector loading. Results are merged and deduplicated.
+- **Memory injection at completion time.** `BuildMemoryHint()` loads
+  relevant memories and formats them as a system prompt section, injected
+  after the knowledge base hint. Context-budget aware with configurable
+  character limit. Supports embedding-based relevance filtering when the
+  user's latest message is available.
+- **Persona memory configuration.** Personas gain `memory_enabled` (bool)
+  and `memory_extraction_prompt` (text) fields. When memory is enabled on
+  a Persona, conversations with that Persona contribute to persona-scoped
+  memory extraction. Custom extraction prompts allow specialization
+  (e.g. helpdesk FAQ extraction vs. tutoring progress tracking).
+- **Database migrations.** Four new migration files (Postgres + SQLite):
+  - `004_v0180_memories.sql` / `sqlite/003_v0180_memories.sql` — `memories`
+    table with composite unique index, full-text search index (GIN/keyword),
+    `memory_extraction_log` tracking table.
+  - `005_v0180_memory_phase2.sql` / `sqlite/004_v0180_memory_phase2.sql` —
+    Persona memory columns, extraction log unique constraint.
+- **API endpoints.** Eight new authenticated endpoints:
+  - `GET /memories` — list user's memories (filterable by status/query)
+  - `GET /memories/count` — active + pending counts
+  - `PUT /memories/:id` — edit a memory's key/value/confidence
+  - `DELETE /memories/:id` — delete a memory
+  - `POST /memories/:id/approve` — approve a pending memory
+  - `POST /memories/:id/reject` — reject (archive) a pending memory
+  - `GET /admin/memories/pending` — admin pending review queue
+  - `POST /admin/memories/bulk-approve` — admin bulk approve
+
+### Changed
+- `CompletionHandler` now accepts an `*knowledge.Embedder` parameter for
+  memory injection with semantic relevance filtering.
+- Persona create/update forms include memory configuration fields
+  (enabled toggle, extraction prompt textarea).
+- Admin settings save handler includes `memory_extraction` and
+  `memory_extraction_enabled` keys.
+- `store.Stores` struct includes `Memories MemoryStore` field.
+- Admin panel sections include Memory with pending review loader.
+
+### Technical Notes
+- Memory embeddings use `vector(3072)` matching the existing KB/notes
+  schema. HNSW index is not used (pgvector limits HNSW to 2000 dims);
+  filtered sequential scans are performant for per-user memory tables.
+  IVFFlat or dimension reduction available as future optimizations.
+- SQLite hybrid recall loads embeddings into Go and computes cosine
+  similarity at the application level, reusing the existing
+  `cosineSimilarity()` function from the knowledge base store.
+- Memory tools use late registration (like KB search and note tools)
+  because they require stores and embedder dependencies initialized
+  in main.go.
+
 ## [0.17.3] — 2026-02-28
 
 ### Added