Add structured memory fields with field-level embeddings

[aurexav]

Context
claude-mem separates observations into narrative, facts, and concepts, which improves retrieval granularity. ELF currently embeds full note text only.

Goal
Improve semantic precision by embedding structured fields and allowing field-specific retrieval.

Scope

1. Data model

• Extend the memory note representation with optional structured fields (examples):
  • summary (short)
  • facts (list of short sentences)
  • concepts (list of short phrases)
• Preserve the existing text field as the canonical human-readable note.

2. Write semantics

• add_event: extractor may populate structured fields. Evidence binding requirements must still hold for stored facts.
• add_note: remain deterministic (no LLM). Structured fields are optional input; omission must not change current behavior.

3. Indexing

• Generate field-level embeddings and index them separately.
• Merge field matches back to a single result per note with explicit explain output showing which fields matched.

Non-goals

• No graph database backend.
• No graph traversal queries.

Related issues

• Graph-lite entities and relations as structured fields: Graph-lite entities and relations as structured memory fields (no graph DB) #27

Testing and evaluation

• Add a small evaluation set focused on fact-like queries.
• Compare precision and false positives before/after using the retrieval harness.

Acceptance criteria

• Field-level retrieval improves precision on fact-like queries without degrading general queries.

References

• https://github.com/thedotmack/claude-mem

[aurexav]

Implemented structured note fields and field-level embeddings.

Changes:

• New tables: memory_note_fields, note_field_embeddings.
• add_note accepts optional structured {summary,facts,concepts}. Facts must be evidence-bound deterministically (substring of note text or source_ref.evidence[].quote when present).
• add_event extractor may emit structured; extracted facts must be supported by evidence quotes.
• Worker embeds structured field texts and stores vectors in note_field_embeddings.
• Search augments candidates with Postgres field-embedding matches and exposes matched structured fields via explain.match.matched_fields.

Docs/tests:

• Design doc: docs/plans/2026-02-10-structured-memory-fields-design.md
• Sample eval dataset: docs/guide/eval-structured-facts-sample.json
• cargo make test-rust: PASS

Commit: fd8cd23

[aurexav]

Follow-up verification and fixes:

• Fixed integration test DB reset to truncate new structured field tables.
• Addressed clippy lints so cargo make lint passes.

Verification (local):

• cargo make fmt-check: PASS
• cargo make lint: PASS
• cargo make test-integration (with ELF_PG_DSN and local Qdrant): PASS
• cargo make e2e (context-misranking-harness): PASS

Commits:

• cbc657f (integration reset fix)
• f40b727 (clippy/lint fixes)