Five tools with complete JSON schemas and clear descriptions. Naming follows verb_noun pattern (mem_store, mem_search, mem_get, mem_delete, mem_list). Descriptions are well-written (80-180 chars) and explain WHAT, WHEN, and WHY. All parameters have types and descriptions. Output schemas are documented in code. However, no tool annotations (readOnlyHint/destructiveHint) are present, and error handling lacks recovery guidance. Parameter descriptions could be more prescriptive about constraints (e.g., key format, top_k bounds). Missing examples of error responses.

• Add tool annotations to MCP tool registration: mark mem_store and mem_delete with destructiveHint=true, mem_search/mem_get/mem_list with readOnlyHint=true. This guides agent planning and prevents accidental misuse.
• Enhance error messages with recovery guidance. E.g., 'Embedding failed: connection timeout. Verify EMBED_HOST is reachable or set EMBED_PROVIDER=local. Retry the operation.' instead of bare exception text.
• Document parameter constraints explicitly: key must be 5-500 chars, alphanumeric + spaces/hyphens; top_k must be 1-100; limit must be 1-1000. Add these to parameter descriptions.
• Add output schema documentation to tool descriptions. E.g., mem_search should state: 'Returns {results: [{key, value, category, similarity, final_score, age_days, pinned}], count}' so agents know what fields to expect.
• Implement idempotency guarantees: document that mem_store with same key+value is idempotent (updates timestamp only), and mem_delete is idempotent (no error if key missing). This enables safe retries.

Inferred effective spec: 2026-07-28+.

• Relies on Stateful initialize / Mcp-Session-Id (removed; protocol is stateless) - make each request self-contained