Summary

The holographic memory provider's FactRetriever._fts_candidates passes the raw user query directly to FTS5's MATCH operator. FTS5 defaults to AND-between-tokens, which means any multi-word prose query requires every token to co-occur in a fact. For the prefetch() path where the query comes straight from the user message, this reduces recall to near-zero on natural-language prompts.

Example, before this fix:

query: "what happened with the deployment rollback"
FTS5 MATCH: "what AND happened AND with AND the AND deployment AND rollback"
results: 0  (nothing has all six tokens)

query: "deployment OR rollback"
results: 5  (normal recall)

The prefetch hook in run_agent.py that injects memory context on every turn was therefore silently missing relevant facts whenever the user phrased their message in prose.

Fix

Add _sanitize_fts_query() that:

• tokenizes and lowercases the query
• drops standard English stopwords and <2-char tokens
• strips FTS5 operator characters from each remaining token
• OR-joins the survivors as phrase literals: "tok1" OR "tok2" OR ...
• falls back to the raw query if nothing survives sanitization (pathological inputs)

No changes to the HRR + Jaccard + trust reranking — those keep precision high once the candidate pool isn't empty.

Test plan

Ships with 10 new tests in tests/plugins/memory/test_holographic_retrieval.py:

• parametrized sanitizer unit tests (stopword drop, single content word, pure-stopword fallback, FTS5-special stripping, empty input)
• FTS5 crash-safety test against problematic inputs (quotes, stars, parens, carets, colons, hyphens, long strings)
• integration tests against an in-memory MemoryStore:
  • natural-language prose query recovers the relevant fact (the exact regression this fix targets)
  • single-keyword query still works
  • pure-stopword query returns [] without crashing

Existing memory/fact-store test suite (329 tests) still passes.

pytest tests/plugins/memory/test_holographic_retrieval.py  # 10 passed
pytest tests/ -k "memory or fact_store or retriev or holographic"  # 329 passed, 4 skipped

Notes

• Pure bug fix, no API surface change, no config change.
• Stopword list is the standard English set (baked in). Could be made configurable later if multi-language is desired, but that's out of scope here.
• The sanitizer is a @classmethod so tests can call it directly without instantiating a retriever + store.