Fact extraction (v0.2.9)

When a memory's content exceeds 500 characters, mnueron can run a Claude Haiku call to extract 1-5 discrete facts — decisions, recommendations, conclusions, action items — and save each as its own searchable memory.

Cost: ~~$0.002 per extraction. Combined with summarization (~~$0.0015) your max per-long-save cost is ~$0.005.

Enabling extraction

Two opt-in paths:

Per-deploy: set environment variable ENABLE_FACT_EXTRACTION=true on your hosted account. Every memory above the threshold gets extraction automatically.

Per-call: include metadata.extract_facts: true in your save:

mem.save(
    content=long_assistant_response,
    namespace="user-123",
    metadata={"extract_facts": True},
)

What gets saved

Each extracted fact becomes its own memory in the same namespace:

Field	Value
`content`	The fact text (1-2 sentences)
`namespace`	Same as parent
`tags`	`["extracted-fact", <category>]`
`source`	`"fact-extraction"`
`source_ref`	parent memory id
`metadata.derived_from`	parent memory id
`metadata.category`	`decision` / `recommendation` / `fact` / `action_item`

To query just the extracted facts later:

# All decisions across this user's history
mem.search(
    "what did we decide about pricing",
    metadata_filter={"category": "decision"},
    namespace=f"user-{user_id}",
    k=5,
)

Failure mode

Fail-open. If the Anthropic API call fails (rate limit, network error, no ANTHROPIC_API_KEY configured), no facts are saved and the parent memory write still succeeds normally. No client-side error.

Trigger timing

Fact extraction runs after the parent memory is saved (so the parent id is known) and doesn't block the response — the user sees the parent insert succeed immediately, and children land within a few seconds as the Haiku call completes.