Server has 13 tools with explicit schemas and descriptions. Naming follows verb_noun pattern (mem_*) consistently. Most tools have parameter descriptions and type definitions. However, descriptions are often terse (10-50 chars) and lack actionable context for LLM selection. Output schemas are not documented, LLMs cannot predict response structure. Error handling is minimal; no recovery guidance. Tool composition is reasonable but some tools conflate concerns (mem_batch, mem_evidence). Security annotations present (read_only_hint, destructive_hint, idempotent_hint) but no permission gates or audit trail guidance.

• Document output schemas for all tools. Example: mem_recall returns {hits: [{id, kind, score, content}], diagnostics?: {arms, contrib, timings}}. Use JSON Schema in tool metadata.
• Expand descriptions to 100-200 chars. mem_status: 'Get server health: store size in MB, message/fact/summary counts, backend status (vector/FTS), and any maintenance errors from last run.'
• Split mem_batch into mem_batch_remember (messages) and mem_batch_fact (facts). Split mem_evidence into mem_cite, mem_compute, mem_conflicts.
• Add error handling to tool descriptions. Example: 'Returns {id, error} on failure. If owner mismatch, error is "Permission denied: fact owned by X". If not found, error is "Fact 123 not found, try mem_recall to search."'
• Add enum constraint to mem_recall 'rel' parameter: {type: 'string', enum: ['supports', 'contradicts', 'supersedes', 'derives_from']}. Same for mem_link.