managed-memory: search-first recall via the new entries:search tool

[kevinyan122]

What

Adds search_memory as a sixth tool to the managed-memory skill and makes it the primary recall path, replacing the previous list_memories → get_memory flow (which was documented as a stopgap until search shipped). Synced to all 5 templates that carry the skill.

• New tool prompt documents the query contract for the lexical (BM25) entries:search API: a natural-language question or keywords both work, phrased with the words you'd expect to appear in the memory itself; top_k default 10, max 50; results come back ranked with full contents inlined, so recall is a single call with no get_memory follow-up.
• get_memory / list_memories demoted to fallbacks (exact-path reads of [has_contents] entries; full-inventory asks like "what do you remember about me?").
• Save dedup now searches first instead of listing.
• MEMORY_INSTRUCTIONS rewritten to match: search-first recall, empty-result semantics (nothing matched those words ≠ nothing stored), raise top_k or fall back to list_memories for very broad questions.
• Limits/troubleshooting: BM25 scores are unbounded (ranking only); newly written entries take a few seconds to become searchable (list/get see them immediately); fallback guidance for workspaces where entries:search hasn't rolled out yet.

Testing

Dogfooded on agent-langgraph-advanced against a staging workspace (store kevinyan.default.kevin_test), locally and deployed to Databricks Apps with OBO scope resolution:

• 19-conversation eval in a fresh scope: all 7 save-side cases clean (dedup, update-in-place on contradiction, no saving of transient chatter); recall probes confirmed single-call search recall, honest misses on unstored facts, list_memories for inventory asks.
• Verified the write→searchable index lag and the unbounded-score behavior documented here.

🤖 Generated with Claude Code