memxt has 10 tools with complete input schemas and descriptions, but quality is inconsistent. Tool names are domain-specific (remember, search, forget, wake_up) rather than verb_noun patterns (store_memory, search_memories). Descriptions are brief (10-80 chars) and lack context on WHEN to use each tool or what happens on error. Parameters lack type constraints (enums, ranges, patterns). Output schemas are not documented, LLMs cannot predict what search() or wake_up() returns. No error handling guidance. The palace/wing/drawer metaphor is creative but adds cognitive load for LLM reasoning.

• Rename tools to verb_noun pattern: remember→store_memory, forget→delete_memory, dream→consolidate_memory, wake_up→generate_continuity_brief, anchors→verify_anchors. This makes intent obvious to LLMs.
• Document output schemas for all tools. E.g., search() returns {results: [{id, content, wing, kind, score}], total_count, next_cursor}. wake_up() returns {brief: string, tokens_used, level_breakdown}.
• Add enum constraints to parameters. Declare search.mode as enum: ['hybrid', 'memories', 'documents', 'facts', 'episodes']. Add min/max to budget (e.g., 100 - 8000 tokens).
• Expand descriptions to 50 - 150 chars with WHEN/WHY context. E.g., 'Consolidate memory by demoting infrequently-accessed facts and clustering related memories. Call this periodically to reduce storage and improve search latency.' Include prerequisites: 'Requires at least 100 stored memories.'
• Add error handling guidance. E.g., forget() should document: 'Returns error if drawer_id not found. Suggest search() to find valid IDs. Deletion is irreversible.' dream() should note: 'dry_run=true simulates consolidation without persisting; use to preview impact before committing.'