Greptile Summary

This PR fixes a bug in hybrid memory search mode where multi-word queries (e.g. Rocky Point bagging) returned 0 results because the full query string was passed to FTS5 as an AND query, requiring all tokens to co-occur in the same chunk. The fix introduces searchKeywordsMulti, which mirrors the FTS-only mode strategy: extract individual keywords, search each in parallel, and merge by highest textScore per chunk id.

Key changes:

• src/memory/manager.ts: New private searchKeywordsMulti method replaces the single searchKeyword(cleaned, candidates) call in hybrid mode. The merging logic (dedup by chunk id, sort by textScore, slice to limit) is correct and consistent with the FTS-only path.
• src/memory/manager.hybrid-keyword-multi.test.ts: Well-constructed regression tests using 8-token chunks and flat embeddings to isolate the keyword path, directly reproducing the reported failure.

Implementation quality: The refactoring maintains consistency with the FTS-only search path: individual searchKeyword calls are protected with .catch() at line 408, while extractKeywords (a pure string tokenization function) is unprotected in both code paths. This is appropriate error handling — FTS errors are gracefully caught, while logic errors surface for investigation.

Confidence Score: 5/5

• Safe to merge — the fix correctly addresses the reported bug with well-tested, consistent error handling.
• The implementation is correct and follows established patterns in the codebase. The searchKeywordsMulti method properly mirrors the FTS-only search strategy by extracting keywords individually, searching each in parallel, and merging by highest textScore. Error handling is appropriate and consistent — FTS errors from individual searchKeyword calls are gracefully caught, while pure string operations behave the same as in FTS-only mode. Regression tests directly validate the fix for the reported multi-word query issue.
• No files require special attention

_{Last reviewed commit: 47aad9c}

💡 Codex Review

https://github.com/openclaw/openclaw/blob/9186827477d361f6ccda5042c864b450d6a1290b/.secrets.baseline#L121-L124
Restore detect-secrets excludes in baseline metadata

This baseline dropped the filters.regex.should_exclude_file / filters.regex.should_exclude_line entries, so detect-secrets scan --baseline .secrets.baseline (the documented local workflow in SECURITY.md) no longer carries the repo’s intended excludes and will start flagging churn from files like pnpm-lock.yaml or known fixture lines whenever they change. CI’s pre-commit hook still passes explicit excludes from .pre-commit-config.yaml, so this mainly regresses local reproducibility and makes baseline regeneration noisy/inconsistent unless those filters are restored in the baseline.

This pull request has been automatically marked as stale due to inactivity.
Please add updates or it will be closed.

Codex review: keeping this open for maintainer follow-up; there is still a little grit to resolve.

Keep this PR open. Current main still does not implement the intended provider-backed hybrid memory-search behavior: it only fans out per keyword in the no-provider FTS-only branch, while provider-backed hybrid search still performs one full-query FTS lookup. The live PR remains open, now has a bounded fan-out repair at head 8ad671b, and maintainer context keeps it as the canonical contributor path after #57711 was closed unmerged.

Best possible solution:

Keep this PR open for maintainer review and landing. The best path is to rebase or port the bounded searchKeywordsMulti implementation into extensions/memory-core/src/memory/manager.ts, keep provider-backed split multi-word regression coverage in extensions/memory-core/src/memory/index.test.ts, preserve the changelog credit, and run the changed gate before merge.

What I checked:

• Current provider-backed hybrid path still uses one full-query FTS lookup: After the no-provider branch, current main computes keywordResults with a single searchKeyword(cleaned, candidates, undefined, sourceFilterList) call when hybrid/FTS are enabled. There is no searchKeywordsMulti helper or HYBRID_KEYWORD_FANOUT_LIMIT in the provider-backed path. (extensions/memory-core/src/memory/manager.ts:428, a820a307dfeb)
• Per-keyword fan-out is limited to the FTS-only branch on current main: Current main extracts keywords and fans out only inside if (!this.provider), after full-query FTS returns no results. That does not cover the provider-backed hybrid case described by this PR. (extensions/memory-core/src/memory/manager.ts:369, a820a307dfeb)
• Full-query FTS still AND-joins tokens: buildFtsQuery tokenizes the raw query, quotes each token, and joins them with AND, so a single multi-word MATCH query can still require all query tokens in the same FTS row/chunk. (extensions/memory-core/src/memory/hybrid.ts:33, a820a307dfeb)
• Current main lacks this regression coverage: Searches for the PR's markers and examples found no current-main implementation or test coverage. The existing hybrid keyword tests in index.test.ts are skipped generic cases, not provider-backed split multi-word regression coverage. (extensions/memory-core/src/memory/index.test.ts:383, a820a307dfeb)
• PR branch implements bounded provider-backed fan-out: The live PR diff adds HYBRID_KEYWORD_FANOUT_LIMIT = 8, adds searchKeywordsMulti/resolveHybridKeywordTerms, replaces the provider-backed single searchKeyword call with searchKeywordsMulti, and adds a provider-backed split multi-word test that asserts MATCH fan-out stays <= 8. (extensions/memory-core/src/memory/manager.ts:62, 8ad671b64b9c)
• Related PR does not supersede this one: Live API data shows feat(memory): enable query expansion in hybrid search mode #57711 is closed and not merged, while this PR is still open. A maintainer comment says ProjectClownfish pushed a narrow repair to this branch so the original contributor path can stay canonical.

Likely related people:

• steipete: Recent current-main history for extensions/memory-core/src/memory/manager.ts and index.test.ts is heavily maintained by steipete, including memory manager/status/search preflight refactors and recent memory CLI/status fixes. The local blame for the inspected current-main search region also routes through the current shallow checkout to Peter Steinberger. (role: current-main memory-core maintainer and recent refactor owner; confidence: high; commits: df65a75f92a3, 9dcd53c0b676, ca27d932b4ff; files: extensions/memory-core/src/memory/manager.ts, extensions/memory-core/src/memory/index.test.ts, extensions/memory-core/src/memory/hybrid.ts)
• vincentkoc: Live PR comments say vincentkoc pushed the narrow ProjectClownfish repair to keep this contributor path canonical. Current-main history also shows vincentkoc touching memory-core manager/tests and related plugin cleanup, making him a strong routing candidate for review and landing. (role: recent memory-core maintainer and PR branch repair author; confidence: high; commits: 8ad671b64b9c, ca26489fe882, 24c39de9c143; files: extensions/memory-core/src/memory/manager.ts, extensions/memory-core/src/memory/index.test.ts, CHANGELOG.md)
• irchelper: Merged PR feat: FTS fallback + query expansion for memory search #18304 introduced FTS fallback plus query expansion for memory search in the older src/memory layout, including manager.ts and query-expansion.ts. This PR is a follow-up to that feature because current main only applies the expansion strategy in the FTS-only path, not provider-backed hybrid search. (role: original merged query-expansion / FTS fallback contributor; confidence: medium; commits: de98782c3736; files: src/memory/manager.ts, src/memory/query-expansion.ts, src/memory/manager-search.ts)

Remaining risk / open question:

• The PR intentionally broadens lexical recall in provider-backed hybrid search, so maintainers should review ranking and precision before landing.
• Per-keyword FTS fan-out increases SQLite MATCH work; the current branch caps fan-out at 8 terms, but representative-index performance sanity is still useful.
• The PR base is older than the checked-out main, so it likely needs a careful rebase around recent memory-core manager and index-test changes.
• Security-sensitive risk appears low in the current diff because it avoids CI, dependencies, lockfiles, package scripts, generated/vendor files, permissions, and secret-handling surfaces.

Codex review notes: model gpt-5.5, reasoning high; reviewed against a820a307dfeb.

ProjectClownfish pushed a narrow repair to this branch so the original contributor path can stay canonical.

Source PR: #39555
Validation: pnpm check:changed
Contributor credit is preserved in the branch history and PR context.

ProjectClownfish could not safely update this branch, so it opened a narrow replacement PR instead.

Replacement PR: #73976
Source PR: #39555
Contributor credit is preserved in the replacement PR body and changelog plan.