Curion provides 2 tools (remember, recall) with basic but incomplete definitions. Both tools have clear verb_noun naming (remember_*, recall_*) and descriptions present, but descriptions are short (under 100 chars) and lack critical context about when/why to use each tool. Input schemas are properly defined with JSON Schema, but descriptions for parameters are minimal. No output schema is documented. Error handling is not visible in the provided code. Tool composition is sound (each does one thing), but lack of structured output documentation and missing error guidance reduce overall quality.

• Expand tool descriptions to 50-200 chars, explaining WHAT each tool does, WHEN the LLM should call it (e.g., 'Call remember after learning new facts about the project to persist them across conversation turns'), and what it enables downstream (e.g., 'Call recall before making decisions to retrieve relevant context').
• Add parameter descriptions that explain constraints: 'text (string, required, 1-10000 chars): The text snippet to store in project memory. Typically a fact, decision, or finding relevant to this project.'
• Document recall output schema: specify whether it returns an array of matched memories, a single best result, or structured objects with relevance scores. Include an example response shape.
• Document remember output schema: specify confirmation format (e.g., {"success": true, "memory_id": "uuid", "stored_at": "ISO8601"}) so LLMs can verify success.
• Add error handling examples: 'If recall finds no matching memories, returns {"matches": [], "error": "No relevant memories found"}. Try a different query or call remember to store initial context.'
• Document recall result limits: 'Returns up to 10 most relevant memories ranked by relevance score. Use limit parameter to adjust (1-100, default 10).' Add limit and offset parameters if not already present.
• Clarify memory scope: Is remember/recall project-local (isolated per project directory), session-local (lost when server restarts), or persistent (backed by SQLite/DB)? State this in both tool descriptions so LLMs understand data lifecycle.

Inferred effective spec: 2025-06-18+.

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