Memkoshi provides 13 tools with basic descriptions and input schemas. However, most descriptions are terse (10-40 chars), parameter descriptions lack detail on constraints/formats, and output schemas are not documented. Tool names follow verb_noun convention (memory_*), but descriptions do not explain WHEN to use each tool or what distinguishes similar ones (e.g., memory_recall vs memory_patterns vs memory_insights). No error handling guidance, no examples of expected output structure, and no documentation of pagination or result limits. Schemas are present but minimal, most parameters lack constraints (enums, ranges, patterns). This is typical of a community server that works but lacks production-grade polish.

• Expand tool descriptions to 80-200 chars. For each, explain: (1) What it does, (2) When to use it vs similar tools, (3) What it returns. Example: 'memory_recall: Search stored memories by keyword. Use this to find relevant past context before making decisions. Returns up to 5 matching memories with relevance scores and timestamps.'
• Document output schemas for all tools. Specify the structure of returned objects, field types, and which fields are always present. Example: 'memory_recall returns {results: [{id, text, category, timestamp, relevance_score}], total_found: int}'.
• Add constraints to numeric parameters. Set min/max for limit (1-100), token_budget (256-16384), priority (1-5). Use enums for categorical params.
• Add error handling guidance. For write tools (memory_commit, memory_approve, memory_reject), document: 'Returns error if memory_id not found (404), call memory_staged() to list valid IDs. Returns error if already approved (409), idempotent retries are safe.'
• Clarify tool relationships. Add a note in memory_recall: 'Use memory_patterns to discover behavioral trends instead of searching by keyword. Use memory_insights for AI-generated summaries of patterns.' This prevents LLM confusion between similar tools.