Smarter Memory Recall: FTS Fallback and Query Expansion in OpenClaw 2026.2.17

One of the quieter but impactful improvements in OpenClaw 2026.2.17 is a significant upgrade to how your agent searches its memory: FTS (full-text search) fallback plus query expansion. If you've ever wondered why your agent sometimes "forgot" something you told it yesterday, this change might solve your problem.

The Problem: Vector Search Has Blind Spots

OpenClaw's memory_search tool uses semantic vector search to find relevant snippets in MEMORY.md and your daily notes. Vector search is great at finding conceptually similar content—but it has weaknesses:

• Exact matches can fail: Searching for a specific name like "ProjectAlpha" might miss an entry that says "Project Alpha" (with a space)
• Short queries struggle: One-word searches don't generate enough semantic signal
• Technical terms get lost: Model names, version numbers, or obscure terms might not have good embeddings

The Solution: Layered Search Strategy

OpenClaw 2026.2.17 introduces a two-layer approach:

1. Query Expansion

Before searching, your query gets expanded with related terms. A search for "AWS deployment" might also look for "Amazon", "cloud", "infrastructure", and "EC2". This catches entries that use different vocabulary for the same concepts.

2. FTS Fallback

If vector search returns low-confidence results, OpenClaw now falls back to traditional full-text search. This is particularly effective for:

• Exact name matches ("@johnsmith", "v2.3.1")
• Rare technical terms the embedding model doesn't understand well
• Very short queries where semantic matching struggles

What This Means for You

More reliable recall: Your agent is less likely to claim it doesn't remember something that's clearly in its memory files.

Better keyword handling: Searching for project codenames, ticket numbers, or specific dates now works more reliably.

No configuration needed: This upgrade works automatically. Your existing memory setup benefits immediately after updating.

Technical Details

The FTS fallback uses the same SQLite FTS5 index that powers other parts of OpenClaw. Query expansion uses a lightweight heuristic approach rather than an LLM call, so it adds negligible latency.

If you want to verify FTS is working, you can check the verbose logs during a memory search—you'll see which layer returned the results.

Try It Out

Update to 2026.2.17 and test with a specific search that previously failed:

/memory_search projectname-v2.3.1

If you had entries with exact matches that vector search missed, they should now surface.

---

GitHub PR: #18304 — Thanks @irchelper for the contribution!

Related: This builds on the Unicode-aware FTS improvements from earlier releases that fixed CJK and non-ASCII query handling.