Mnemosyne has complete input/output schemas and basic descriptions for all 12 tools, but falls short on description quality and parameter guidance. Tool descriptions are minimal (10-30 chars), lacking context on WHEN to use each tool or what distinguishes similar operations (e.g., message_get_latest_unreaded vs message_get_oldest_unreaded). Parameter descriptions are present but generic ('Agent name of the AI instance' repeated 8+ times). No enum constraints, no format guidance, no error recovery hints. Output schemas are defined but not documented in tool descriptions. No tool annotations (readOnlyHint/destructiveHint) despite clear READ/WRITE semantics. Naming is verb-first and mostly clear, but lacks natural-identifier support (agent_name is opaque string, no email/username fallback). No pagination support on list-like tools (messages_get_unreaded, agents_get_details). Security: no visible input validation, no rate limiting, no audit logging in handler code.

• Expand tool descriptions to 100-200 chars. For each tool, answer: What does it do? When should the LLM call it instead of a similar tool? What does it return? Example: 'state_save: Persist an agent's internal state (memory, context, reasoning) for later retrieval. Use this after significant reasoning steps to enable recovery if the agent crashes. Returns the saved state ID and timestamp.'
• Add format and constraint guidance to parameter descriptions. Replace 'Agent name of the AI instance' with 'Unique identifier for the agent (alphanumeric, 3-50 chars, e.g., agent_001). Must match an agent created via agent_create().'
• Add tool annotations: mark state_save, message_create, message_set_as_readed, agent_create as destructiveHint=true; mark all GET tools as readOnlyHint=true. This enables agents to reason about retry safety.
• Implement pagination on messages_get_unreaded and agents_get_details: add limit (default 20, max 100) and offset parameters. Return total_count and has_more in responses. Document the limit in tool descriptions.
• Add input validation and error recovery hints in handler code. Validate agent_name format, return 'Agent not found: agent_xyz. Available agents: [list]. Try agent_create() to register a new agent.' instead of raw exceptions.

Inferred effective spec: 2025-06-18+.

• Relies on Sampling (deprecated) - integrate directly with the LLM provider API