Heimdall has well-structured tool definitions with consistent naming patterns (verb_noun: store_memory, recall_memories, session_lessons, memory_status, delete_memory, delete_memories_by_tags). All 6 tools have descriptions and documented input schemas. However, descriptions are functional but lack guidance on when/why to use each tool, parameter descriptions are minimal, output schemas are not documented, and error handling guidance is absent. The server follows good JSON Schema structure with types and constraints (enums, min/max), but descriptions average ~80-120 chars and are more technical than LLM-optimized. No tool annotations (readOnlyHint, destructiveHint, idempotentHint) despite two destructive tools. The memory system is domain-specific and well-reasoned, but tool definitions could better support agent decision-making.

• Add output schema documentation for each tool. E.g., 'Returns {success: bool, memory_id: string, stored_at: string (ISO 8601)}' so LLMs know what to extract and pass downstream.
• Add tool annotations to source code. Mark delete_memory and delete_memories_by_tags with destructiveHint=true; mark recall_memories and memory_status with readOnlyHint=true; mark store_memory with idempotentHint=true (if idempotent by ID).
• Expand parameter descriptions with decision guidance. E.g., for hierarchy_level, add: 'Use 0 for abstract concepts (e.g., "LLMs process tokens"), 1 for contextual knowledge (e.g., "This project uses FastAPI"), 2 for specific events (e.g., "User reported error on 2024-11-15").'
• Add error recovery hints to tool descriptions. E.g., 'If query has no results, try broadening tags or removing filters. If memory_status shows capacity exceeded, use delete_memories_by_tags to archive old memories.'
• Explicitly recommend dry-run for destructive tools. E.g., 'Always pass dry_run=true first to preview what will be deleted, then execute with dry_run=false if preview is correct.'
• Clarify memory_status purpose with examples. E.g., 'Call this to check total memories stored, storage capacity, and last sync time. Use before intensive store_memory calls to avoid quota errors.'
• Document tool ordering and dependencies in descriptions. E.g., 'Call memory_status first if you suspect capacity issues; then use store_memory or delete_memories_by_tags as needed.'
• Add per-parameter constraints in descriptions where schema alone is insufficient. E.g., for tags array, add: 'Use 1-5 tags per memory; tags should be lowercase, hyphen-separated (e.g., ml-training, bug-fix).'

Inferred effective spec: <=2025-11-25.

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