fix(memory): efficiency and correctness overhaul for extended + legacy memory

[jkyberneees]

Consolidated fix for 21 bugs found in a memory-system audit (5 parallel audit agents, top findings re-verified against the code before fixing). All fixes include regression tests.

Correctness

Extended memory

• Cap bypass on the scan-reject path (regression from fix(memory+guard): quarantine scan-rejected atoms and fix socket_path transport #87): scan-rejected atoms were quarantined before enforceCap ran, so a rejection storm could grow quarantine past max_size_mb and wedge the store (the evictor only evicts trusted atoms). Cap is now enforced on both quarantine paths.
• ReturnAfterBreak summarized the 5 OLDEST atoms — backwards iteration over a newest-first list.
• Final recall ordering discarded relevance: the union of literal + predicted results was re-sorted by pure RetentionScore (degenerating to sort-by-recency with extractor confidence defaulting to 1.0), throwing away the blended 0.6·sim + 0.4·retention score and the paid LLM rerank order. The union is now deduped by ID and sorted by composite score (docs/EXTENDED_MEMORY.md updated accordingly).
• No dedup of extracted atoms — re-stated facts accumulated as duplicates. Adds now dedup by normalized text (refresh CreatedAt, keep higher confidence and original ID).
• Eviction now removes association links; ForgetAtom works for quarantine-only atoms; atom truncation is rune-safe; quarantine TTL eviction no longer writes under a read lock; background user-model inference can't overlap itself or append duplicate pending reviews.
• Fenced/preamble LLM JSON (```json wrappers) made extraction/inference/prediction silently produce nothing — a shared extractJSON helper now strips fences and extracts the first balanced JSON span.

Agent loop

• Skill loader and episode recall re-ran on every loop iteration (up to 300×/turn) whenever the result was empty — the dedup key was only assigned on non-empty results.
• Per-message dedup keys were never reset between Run/RunWithMessages — a REPL user sending the same text twice got no memory hooks the second time.

Legacy memory

• merge_on_write dropped facts when two entries shared a 30-char prefix (substring Replace → ambiguous) — new index-based FactStore.ReplaceAt.
• Episode Search query cache was never invalidated on write/promote/prune — promoted episodes stayed invisible to repeat searches.
• BuildSystemPrompt skipped the extended user-state block when legacy facts were empty, and re-read fact files every iteration because the empty result was never cached.
• The LLM ranker's explicit "none relevant" was overridden with unranked candidates; Consolidate bypassed the facts char cap (and persisted untrimmed entries).

Efficiency

• Index rebuild storm: every per-turn atom add triggered a full vector-index rebuild (dirty flag per add) — now one rebuild per turn via a batch add path (single markDirty, single assoc.Persist).
• Cold embedding cache per rebuild: index/episode rebuilds constructed a fresh embedder each time, so the HTTP embedder's cache never warmed and the entire corpus was re-embedded over the network per atom. The index now reuses one embedder; episode dedup + index share one via new embedding.Shared (HTTP backend only — RP is corpus-fitted and stays per-corpus).
• Recall fan-out: one recall cost ~8 LLM calls (duplicate searches + reranks per predicted intent under a shared 5s budget) — now 2 (rerank only the literal query; type filtering over the existing candidate set).
• Per-candidate file re-parsing in recall — one store load per query now.
• Per-turn extraction blocked the ReAct loop (anaphora + 30s LLM call inline before the next think) — now async via MemoryManager.RunBackground (context.WithoutCancel), keeping the loop responsive.
• Episode extraction died at CLI exit (fire-and-forget goroutine before os.Exit) — session-end work is now tracked and drained with a bounded 15s wait in Agent.Close, so episodes survive odek run/continue/REPL exit.

API additions (no signature changes)

MemoryManager.RunBackground, MemoryManager.WaitForBackground, FactStore.ReplaceAt, embedding.Shared. NewLLMRanker's "none" case returns an unexported sentinel error.

Verification

• 30+ new regression tests (batch rebuild count, embedder reuse, composite ordering, cap enforcement on reject path, dedup keys, async extraction, CLI-exit drain, merge prefix collision, cache invalidation, fenced JSON, …)
• go test ./... — 27 packages, all green
• go test -race ./internal/memory/... ./internal/loop/... ./internal/embedding/... — green
• golangci-lint run ./... — 0 issues

[cloudflare-workers-and-pages]

Deploying with    Cloudflare Workers

The latest updates on your project. Learn more about integrating Git with Workers.

Status	Name	Latest Commit	Preview URL	Updated (UTC)
✅ Deployment successful! View logs	odek	1931bfe	Commit Preview URL Branch Preview URL	Jul 22 2026, 09:49 AM