feat(memory): rebuild agent memory on one Postgres fact store

[eastriverlee]

What to look at

Seven commits, in order:

1. c55737db the store: migration 030 (memory_episode, memory_fact, memory_profile, memory_job, pg_trgm, and a memory_fact_embedding side table created only where pgvector exists), the repositories, the hybrid search (pgvector cosine + pg_trgm word similarity, reciprocal rank fusion), and the job outbox with FOR UPDATE SKIP LOCKED claims.
2. 0daf5dd4 ingest: one pipeline for what the agent writes (memory_remember) and what it did (a finished task run). The runtime embeds the episode, offers the nearest live facts the requester may read as candidates, and a low-tier model returns facts related as new / supersedes / reinforces. The runtime rejects any relation to a fact it did not offer. Three tools for the model: memory_search, memory_remember {content}, memory_forget {factIDs} (IDs only from a search in the same task).
3. b6bc9cfe removal: Graphiti sidecar, Kuzu, MEMORY.md, the in-process queue, admin graph endpoints, health probe, config; migration 031 drops the graphiti tables; the namespace list becomes one memory.SecurityLabel per launch.
4. 7c408aca admin: GET /admin/api/memory/facts (a person's profile and the live facts they may read) and POST /admin/api/memory/facts/forget, for the device web page.
5. 922d1abb fix: lexical search when the vector table is absent (CI runs plain Postgres).
6. 27326c18 the store moves to its own module, bluememo, vendored at .dependency/bluememo like the harness; internal/memory keeps only the host adapters. Circles nest: a circle fact names one or more circles, memberCircles on a circle policy says which circles belong to it, and a member of a containing circle reads the contained circle's facts, transitively. Writing to a circle still needs direct membership. migrations/030_memory_store.sql is a verbatim copy of bluememo's schema with a drift test.
7. d45e13bc no company-wide scope: a fact with no circles is read by its owner alone; a fact that names circles is read by their members (and containing circles) too; sharing with everyone is sharing with the member circle. memory_search returns each fact's circles instead of a scope.

Design: docs/internal/memory-redesign.md in the internkim repository (arrives with the submodule bump, internkim PR feat/memory-store).

Evidence

• go vet ./... and go test ./... green; catalogs regenerated. bluememo's own suite (unit + Postgres on pgvector/pgvector:pg16 and plain postgres:16) green, covering multi-circle facts, transitive containment with a cycle, and the reader predicate in SQL and Go saying the same thing.
• Postgres integration (BLUECLAW_TEST_POSTGRES_URL, run on pgvector/pgvector:pg16 and on plain postgres:16): reader filter across scope, rank and classes; supersede hides the old row and keeps it; expiry; scoped forget; vector ranking; job dedupe/claim/lease/retry; a finished task run observed through the real transition observer ends as rows in all three tables and Recall returns them.
• Live model check (BLUECLAW_LIVE_LLM_TEST=1, real low-tier model + real embedding model): a Korean transcript yields a preference and a temporary fact with its expiry; a correcting transcript supersedes the team fact and reinforces the preference; recall ranks the corrected fact first; a profile builds. It passed once on 2026-09-02 after fixing a ranking fault it found (a multiplier for reinforced preferences flipped adjacent reciprocal ranks; reinforcement is now a tie-breaker). Not run in CI.
• Embedding benchmark (BLUECLAW_EMBEDDING_BENCHMARK=1, 30 Korean facts, 20 hand-marked queries, 1,024 dims): qwen3-embedding-8b, qwen3-embedding-4b and pplx-embed-v1-4b all at R@1 1.000; pplx-embed-v1-0.6b 0.950; bge-m3 R@1 0.950 and R@3 1.000. bluememo's library default is baai/bge-m3; the host names the model it actually uses (a device: bge-m3 from its own llama.cpp; a direct-execution host: pplx-embed-v1-4b on OpenRouter). Vectors from two models do not compare, so search ranks vectors only from the store's own model and a reembed job, enqueued at every worker start, moves live facts onto it; the memory_guided_followup fleet scenario now calls memory_remember and requires the remembered sentence in the next task's model context. The live model check passed again against bge-m3 on 2026-09-03.

What the model now sees

memory_search 217 bytes, memory_remember 428 bytes, memory_forget 179 bytes of tool description (was 599 bytes for two tools). The ingest and profile instructions are sent only on the background low-tier calls, never in the agent turn.

Not done here

• pgvector on the bookworm guest rootfs and the capabilityd remote embedding default are in the internkim PR.
• Import of existing MEMORY.md files: none exist on a device this ships to first; the design keeps the import job kind for when one does.

🤖 Generated with Claude Code