Two tools with minimal descriptions and incomplete schemas. 'recall' has a reasonable 89-char description and three typed parameters (query, limit, min_score), but lacks output schema documentation. 'memory_context' has only a 95-char description and zero input parameters, making it difficult for LLMs to understand when to invoke it or what it returns. Neither tool has error handling guidance, recovery hints, or composition context. Parameter descriptions are present but generic. No tool annotations (readOnlyHint, destructiveHint, idempotentHint) despite both being READ_ONLY. Output schemas are completely undocumented, LLMs cannot plan downstream calls or extract structured data.

• Document output schema for 'recall': return type (array of objects), field names (id, text, score, source, layer), types, and example. State result limit (e.g. 'Returns up to 50 results').
• Add explicit output schema for 'memory_context': what fields comprise the 'metadata snapshot'? (e.g. fact_count, semantic_count, document_count, last_updated). Include types and example.
• Add readOnlyHint: true to both tools' definitions in server.py to signal they are safe to retry and have no side effects.
• Expand 'min_score' parameter description: 'Minimum relevance score threshold (0.0 - 1.0, default 0.3). Lower values return more results but with lower confidence.'
• Add error handling guidance: 'If no results found, try broadening the query or lowering min_score. If timeout occurs, reduce limit or try a simpler query.'