PR #81423: fix(memory): preserve qmd lexical search for hyphenated queries

• Repository: openclaw/openclaw
• Author: giodl73-repo
• State: open | merged: False
• Link: https://github.com/openclaw/openclaw/pull/81423

Description (problem / solution / changelog)

Summary

• Preserve the raw user query for QMD lexical (lex) searches so dashed identifiers, dates, and version strings can still match exactly.
• Normalize only QMD semantic sub-searches (vec / hyde) by replacing word-internal hyphens with spaces before sending the v2 qmd.query searches array.
• Add focused coverage for hybrid and vector-only QMD MCP payloads.

Fixes #81328.

Tests

• node scripts/test-projects.mjs extensions/memory-core/src/memory/qmd-manager.test.ts --reporter=verbose -t "hyphenated tokens|vector-only QMD"
• node scripts/test-projects.mjs extensions/memory-core/src/memory/search-manager.test.ts --reporter=verbose -t "falls back to builtin search when qmd fails with sqlite busy|keeps original qmd error when fallback manager initialization fails"
• pnpm check:changed
• git diff --check

Full qmd-manager.test.ts run on Windows reached 104/105 passing; the remaining failure was the existing symlink-permission case (EPERM creating a symlink), unrelated to this change.

Real behavior proof

Behavior addressed: memory_search should not lose QMD lexical recall just because a query contains hyphenated tokens such as sqlite-vec, 2026-05-04, or multi-agent.

Real environment tested: Local Windows source checkout at C:\src\openclaw-pr-81328 running the OpenClaw test harness for the QMD manager. The local machine does not have the qmd binary installed, so this draft has source-level/runtime-boundary proof rather than a live QMD daemon proof.

Exact steps or command run after this patch: Ran the focused QMD manager regression tests for the v2 qmd.query payload construction and the focused fallback-manager checks listed above.

Evidence after fix: The focused QMD manager test captures the exact mcporter call qmd.query arguments and verifies this payload shape for a dashed query:

{
  "searches": [
    { "type": "lex", "query": "sqlite-vec backend health 2026-05-04 multi-agent" },
    { "type": "vec", "query": "sqlite vec backend health 2026 05 04 multi agent" },
    { "type": "hyde", "query": "sqlite vec backend health 2026 05 04 multi agent" }
  ]
}

The vector-only regression verifies vsearch also sends the semantic-safe query:

{
  "searches": [
    { "type": "vec", "query": "sqlite vec backend health" }
  ]
}

Observed result after fix: The lexical branch keeps the raw query, so QMD can still match dashed identifiers and dates. The semantic branches no longer receive word-internal hyphens that QMD v2.0.1 treats as NOT-operator syntax, avoiding the validation failure that forced OpenClaw to fall back to the builtin index.

What was not tested: I did not run a live QMD 2.0.1 daemon or a full memory_search command against an installed QMD index in this environment. The PR is draft until we either get live QMD proof or a maintainer accepts the focused boundary test proof.

Changed files

• CHANGELOG.md (modified, +1/-0)
• extensions/memory-core/src/memory/qmd-manager.test.ts (modified, +74/-0)
• extensions/memory-core/src/memory/qmd-manager.ts (modified, +8/-3)