Server implements 9 memory management tools with complete JSON schemas and descriptions. All tools have proper input schemas with type definitions and parameter descriptions. However, descriptions are generic and lack LLM-optimized guidance on WHEN to use each tool vs. alternatives. No output schemas documented. Error handling is minimal, no recovery guidance or actionable error messages. Tool composition is reasonable (single responsibility per tool), but descriptions don't explain dependencies or multi-step workflows. Naming follows verb_noun convention consistently (memory_save, memory_get, memory_search, etc.), which is good. Parameter descriptions are present but lack constraints (e.g., no max length for 'key', no guidance on category enum values). No tool annotations (readOnlyHint, destructiveHint) despite clear risk levels (WRITE, DESTRUCTIVE). Missing: output schema documentation, error recovery guidance, parameter constraints, and tool annotations.

• Add output schemas to all tools. Document the structure of memory objects returned by memory_search, memory_list, etc. Include fields like 'id', 'key', 'content', 'category', 'tags', 'importance', 'created_at' so LLMs know what to extract.
• Rewrite descriptions to be LLM-optimized (50-200 chars). Example: 'memory_get: Retrieve a single memory by key. Use when you know the exact key; use memory_search to find memories by content.' This clarifies intent and distinguishes from similar tools.
• Add tool annotations: mark memory_save with 'idempotentHint: true' (same key overwrites), memory_delete and memory_delete_by_category with 'destructiveHint: true' to warn LLMs of irreversible side effects.
• Add parameter constraints: 'key' max 255 chars, 'category' enum ['preferences', 'facts', 'decisions', 'context', 'general'], 'limit' max 100, 'importance' range 1-5 (already present). Document these in descriptions, not just schema.
• Add error recovery guidance in descriptions. E.g., 'memory_get: If key not found, try memory_search() with partial content. Returns null if not found.' This guides LLM's fallback strategy.

Inferred effective spec: <=2025-11-25.

• Relies on Logging (deprecated) - log to stderr or use OpenTelemetry