Engram MCP server has 7 well-named tools with consistent verb_noun patterns (search_memories, add_memory, etc.). All tools have descriptions (avg ~60 chars, baseline 194 chars) and input schemas with typed parameters. However, descriptions are terse and lack context for LLM selection. Parameter descriptions are minimal (e.g., 'Search query string' vs 'Natural language query for semantic memory search'). Output schemas are not documented in the source, only inferred from handler implementations. Error handling is present but generic. No tool annotations (readOnlyHint, destructiveHint) despite clear risk classifications. Missing dependency hints and recovery guidance.

• Expand tool descriptions to 100-200 chars. Include WHEN to use (e.g., 'Use search_memories for semantic lookup by topic; use list_memories to browse all memories in order'). Add state-modification warnings for add_memory, update_memory, delete_memory.
• Document output schemas explicitly in tool definitions. Specify response fields (e.g., search_memories returns {results: [{id, content, similarity_score}], total_count}). This enables LLMs to chain tools correctly.
• Add tool annotations: mark delete_memory with destructiveHint=true, search_memories/list_memories/get_retrieval_log with readOnlyHint=true. Agents will treat them appropriately.
• Add parameter constraints in descriptions: 'limit (1-100, default 20)', 'threshold (0.0-1.0, default 0.5)'. Include validation error messages in tool responses (e.g., 'Invalid threshold: must be 0.0-1.0').
• Implement confirmation pattern for delete_memory: add optional 'confirm' boolean parameter. Return {status: 'confirmation_required', memory_id, content_preview} if confirm=false, then execute on confirm=true.