Contextwell demonstrates solid tool design with complete input schemas, clear descriptions, and thoughtful parameter organization. All 6 tools have explicit descriptions (194 - 450 chars, within baseline 34 - 392 range). Parameters are well-typed with enums for constrained inputs (type, scope, format). However, output schemas are not documented, LLMs cannot predict response structure for chaining. Error handling is minimal; most tools return plain strings rather than structured error objects with recovery guidance. No tool annotations (readOnlyHint, destructiveHint) despite clear risk levels (WRITE, DESTRUCTIVE, READ_ONLY). Naming is verb-forward and clear (remember, recall, forget, list_memories, export_memories, import_file). Parameter descriptions are detailed and include format hints (ISO 8601, git repo context). Missing: output schema documentation, structured error responses, tool annotations.

• Document output schemas for all tools. Example for recall: 'Returns array of {id, content, type, scope, tags, created_at, score (if include_score=true), chunk_of (if chunked)}'. This enables LLM chaining.
• Add tool annotations via FastMCP: @mcp.tool(annotations={'readOnlyHint': True}) for recall/list_memories/export_memories; @mcp.tool(annotations={'destructiveHint': True}) for forget; @mcp.tool(annotations={'idempotentHint': True}) for remember (if idempotent with allow_duplicate=false).
• Wrap error responses in structured objects: {'error': 'invalid_expires_at', 'message': '...', 'recovery': 'Pass a valid ISO 8601 datetime or leave empty'}. This guides LLM recovery.
• Document pagination behavior: 'recall returns up to k results. If results are truncated, call again with adjusted query or higher k. Total count not returned; use list_memories for full inventory.'
• Add optional dry_run parameter to forget: 'When true, returns what would be deleted without committing. Agents can preview before confirming with dry_run=false.'