Summary

• Problem: memory_search returns 0 hits for multi-word queries (e.g. Rocky Point bagging, retail garden products) in hybrid mode, even when the terms appear 17+ times across indexed files.
• Why it matters: Users relying on memory search for domain-specific multi-word terms get silently empty results with no indication of why, making the feature appear broken.
• What changed: Hybrid mode now searches each extracted keyword individually and merges results by highest score (OR-style recall), matching the strategy already used by FTS-only mode.
• What did NOT change (scope boundary): buildFtsQuery, searchKeyword, the FTS-only code path, vector search, scoring weights, and all other memory subsystems are untouched.
• Note that test cases were written with assists of AI.

Change Type (select all)

• Bug fix
• Feature
• Refactor
• Docs
• Security hardening
• Chore/infra

Scope (select all touched areas)

• Gateway / orchestration
• Skills / tool execution
• Auth / tokens
• Memory / storage
• Integrations
• API / contracts
• UI / DX
• CI/CD / infra

Linked Issue/PR

• Closes memory_search returns 0 hits for specific terms despite indexed content openclaw/openclaw#39484

User-visible / Behavior Changes

Multi-word memory_search queries in hybrid mode now return results when the query terms span separate chunks. Previously these returned 0 hits. Single-word queries and FTS-only mode behavior are unchanged.

Security Impact (required)

• New permissions/capabilities? No
• Secrets/tokens handling changed? No
• New/changed network calls? No
• Command/tool execution surface changed? No
• Data access scope changed? No

Repro + Verification

Environment

• OS: Linux (WSL2 / native)
• Runtime/container: Node 22+
• Model/provider: any (bug is in FTS path, provider-independent)
• Integration/channel (if any): N/A
• Relevant config (redacted): memorySearch.query.hybrid.enabled: true (default)

Steps

1. Index a document where query terms appear in separate sections (e.g. "Rocky Point" in one paragraph, "bagging" in another).
2. Run memory_search with query Rocky Point bagging.
3. Observe result count.

Expected

• At least one result returned referencing the indexed document.

Actual

• Before fix: 0 results returned despite 17+ term instances across indexed files.
• After fix: results returned correctly.

Evidence

• Failing test/log before + passing after

src/memory/manager.hybrid-keyword-multi.test.ts — 2 tests fail against the reverted manager.ts (0 results asserted > 0), both pass with the fix applied.

Human Verification (required)

• Verified scenarios: Multi-word cross-chunk query (Rocky Point bagging); two-word queries (bagging operations, retail garden products); single-word queries unaffected; FTS-only mode unaffected.
• Edge cases checked: Reverted manager.ts confirms tests correctly catch the regression; restored fix confirms green.
• What you did not verify: Live index with real embedding provider; very large documents with hundreds of chunks; non-English keyword extraction paths.

Compatibility / Migration

• Backward compatible? Yes
• Config/env changes? No
• Migration needed? No

Failure Recovery (if this breaks)

• How to disable/revert this change quickly: revert src/memory/manager.ts to parent commit (3eec79bd6).
• Files/config to restore: src/memory/manager.ts only.
• Known bad symptoms reviewers should watch for: hybrid keyword search returning unexpected results for single-word queries (should not occur — single-keyword path is identical before and after).

Risks and Mitigations

• Risk: per-keyword FTS searches increase the number of SQLite MATCH queries from 1 to N per hybrid search call.
• Mitigation: extractKeywords already strips stop words and short tokens, keeping N small for typical queries (2–5 terms). The candidates multiplier already caps each individual lookup at 200 rows. No change for single-word queries.