Module extraction 

The provider-agnostic post-turn memory extractor (docs/reference/personas.md §5, issue #214).

After a turn commits, a cheap model distills the transcript into durable facts about the person — never in the turn’s hot path. The extractor is handed the persona’s existing ACTIVE facts so contradictions come back as invalidations (paired with the replacing fact) rather than duplicates; the store then closes the old fact’s validity interval — invalidate, never delete.

Like the participation classifier next door, this is deliberately provider-agnostic and tolerant on the way out: a reply that isn’t the expected JSON extracts nothing (memory is best-effort enrichment; a flaky cheap model must never break the pipeline).

Structs

CandidateFact

One fact the extractor proposes to remember.

ExistingFact

An existing active fact, as shown to the extractor for contradiction checks.

ExtractedMemories

What one extraction pass proposes.

Invalidation

One existing fact the extractor proposes to invalidate (contradicted by this turn).

Constants

MAX_FACTS_PER_TURN

Most facts one turn may add. A chatty turn distills to a few durable facts; dozens means the model is transcribing, not distilling.

MIN_CONFIDENCE_BPS

Write-time confidence floor (#796).

Functions

extract_memories

Distill one committed turn into memory operations.

looks_like_authorization_claim

The post-parse authorization-claim heuristic (#1925).

looks_like_pii

The post-parse PII heuristic (#796): a refused-category keyword or a long digit run anywhere in the text.