Server has 5 tools with complete schemas and descriptions. Naming follows verb_noun convention (search_memory, get_conversation, list_conversations, ingest_exports, memory_stats). Descriptions are present but brief (avg 95 chars, below 194-char baseline). All tools have input schemas with type definitions. However, output schemas are not documented, LLMs cannot predict response structure. Parameters lack detailed constraints (e.g., limit ranges, format specs). Error handling is minimal; no recovery guidance visible. No tool annotations (readOnlyHint/destructiveHint) despite clear read/write distinction.

• Document output schemas for all tools. Define response structure as JSON Schema (e.g., search_memory returns {results: [{id, title, source, score, snippet}], total_count, next_cursor}). This enables LLMs to plan downstream calls and extract fields correctly.
• Add tool annotations: mark search_memory, get_conversation, list_conversations, memory_stats with readOnlyHint=true; mark ingest_exports with destructiveHint=true. Agents will know which tools are safe to retry.
• Add parameter constraints: search_memory query (min 1 char), limit (1 - 100, default 10); list_conversations limit (1 - 100, default 50), offset (≥0); ingest_exports sources (array of 'claude-desktop'|'chatgpt'|'claude-code'). Include these in descriptions.
• Enhance error messages with recovery hints. E.g., 'Conversation not found: {conv_id}. Try search_memory(query=...) to find conversations by content, or list_conversations() to browse all.'
• Expand memory_stats description to ~100 chars: 'Get statistics about the memory database: total conversations, messages, tokens indexed, database size, last ingestion time. Use this to monitor memory health and plan ingestion schedules.'