RECALL presents a well-intentioned memory system for agents with 15 tools spanning search, retrieval, and recording functions. However, the evaluation reveals critical gaps in parameter descriptions, missing output schemas, and insufficient error handling guidance. While tool names are action-oriented (memory_hybrid_search, add_breadcrumb, get_recent_decisions) and follow verb_noun convention, the rubric's hard constraints apply: many parameters lack descriptions or detail, and output structures are not documented. The server uses STDIO transport only, which triggers a hard cap of 50 on protocol readiness. The codebase shows structured implementation but does not expose complete schema documentation in accessible form. Most tools fall in the 50-65 range individually; averaging to 58 overall.

• Document complete output schemas for all 15 tools. Example for get_decision: {"id": number, "decision": string, "category": string, "reasoning": string, "confidence": 'high'|'medium'|'low', "status": 'active'|'superseded'|'reverted', "created_at": ISO8601, "updated_at": ISO8601}. This enables agents to chain results and extract required fields.
• Expand all parameter descriptions to 50 - 150 characters with explicit constraints. Example: 'importance (1 - 10): numeric priority level; 1=low, 10=critical. Controls ranking in search results.' Instead of: 'Importance level 1-10'.
• Add formal enum constraints to category and confidence parameters using JSON Schema enum arrays. Example: {"category": {"type": "string", "enum": ["context", "note", "todo", "reference"]}}.
• Define tool annotations for all write operations. Mark add_breadcrumb, add_decision, add_learning with 'destructiveHint' = true (or add_note_like = false if non-destructive). Mark all get_* tools with 'readOnlyHint' = true. This signals retry safety to agents.
• Add pagination: get_recent_breadcrumbs should return {"breadcrumbs": [...], "total_count": number, "limit": number, "offset": number}. Enables agents to fetch in batches and avoid context explosion.
• Clarify LoA terminology. Either rename get_loa_entry to get_lesson_of_arms_entry (explicit) or expand tool description: 'Retrieve a specific LoA (Lesson of Arms, a record of combat decisions and tactics) entry by ID. Use this to track historical decision patterns.'
• Add error guidance: 'If importance is out of range (1 - 10), the tool will return error_code: INVALID_IMPORTANCE with message: "Importance must be 1 - 10; got {value}. Valid values: low(1 - 3), medium(4 - 6), high(7 - 8), critical(9 - 10)."'