feat(memory): add an opt-in entity extraction model - #170

Stacked on #169: the first commit is that PR. Review 47a2204.

What

An opt-in entity extraction model for remember and import.

• internal/memory/extract is a client configured by MNEMON_EXTRACT_MODEL, MNEMON_EXTRACT_ENDPOINT (the full URL, for example Workers AI .../ai/run/@cf/cloudflare/clairic), MNEMON_EXTRACT_API_KEY, and MNEMON_EXTRACT_TIMEOUT. It sends the memory as state with one list field and accepts both the Workers AI envelope and a bare System One response.
• Model output is treated as untrusted. An item survives only if, after trimming edge quotes and clause punctuation, it has a letter, at least two runes, at most 200 bytes, and an occurrence in the memory that is a whole token rather than a fragment (TW in zh-TW, 5.1 in 5.1.0, w/demo in ~/w/demo). Duplicates and items past 50 are dropped, matching the --entities limits.
• graph.ResolveEntitiesWithModel uses the model's entities only when resolution relies on extraction alone: merge mode with nothing provided, or auto mode. They replace the all-caps acronym guess, while CamelCase, file paths, URLs, @mentions, title marks, the dictionary, and the known-entity index still apply.
• The request runs before the write transaction, beside the embedding. A failure prints one warning and falls back to heuristic extraction, and import stops asking after the first failure.
• Memories stored with entities, and recall queries, never call the model.

Why

The regex finds identifiers, paths, and acronyms but misses most names: multi-word people and organizations, and single-segment names on first mention. Agent hosts usually pass entities, but hooks, imports, and other writers often do not. An extraction model returns spans of the input in one input-only pass (about 0.5 to 0.9 s per memory with Clairic, billed only for input tokens), so it fits a write path without a generative LLM.

The exact regex paths stay because, in probes on real memories, the model split paths (~/ and w/...), cut versions, and left trailing brackets, which are the cases those patterns match exactly. Clairic's list-level confidence came back near 1e-12 on every call, so the client does not threshold on it.

Effect

Live remember against Clairic on Workers AI, into a throwaway store:

• A prose memory about people and organizations went from four acronym guesses and one lowercase word to the full names of the people and organizations it mentions.
• A dense technical note kept its file path and identifiers and gained model names and config keys. The model also kept a few ordinary words that pass the whole-token check.
• A bad API key fell back to the heuristic entities with one warning line.

Privacy

Off unless MNEMON_EXTRACT_MODEL is set. When it is set, the full text of each memory stored without entities goes to the endpoint; README and SECURITY.md say so.

Validation

• internal/memory/extract: request shape, both response shapes, whole-token grounding, the 50-item cap, a null list, and errors that name the failure without the key or the content.
• internal/memory/graph: resolution per entity mode, nil versus empty model output, and provided entities bypassing the model.
• cmd/memory: remember with a working, skipped, and failing model; import asking only for drafts without entities and stopping after the first failure.
• gofmt -l is clean and make test (including go vet ./...) passes.

Relation to #161

Independent of the decision-model judge in #161. Both add rows to the same env tables and a package to the Makefile list and architecture graph, so whichever merges second needs a small rebase. The CLI tests use their own stderr helper so the two PRs do not redeclare one.

Release note

Set MNEMON_EXTRACT_MODEL (for example Cloudflare Clairic) to have remember and import extract entities from memories stored without them.