paxm exposes 4 memory-management tools with complete JSON schemas and reasonable descriptions. Tool names follow verb_noun convention (paxm_recall, paxm_remember, paxm_history, paxm_config_doctor). All tools have descriptions (50 - 150 chars, within baseline 34 - 392 range). Input schemas are present with typed properties and descriptions. However, descriptions lack actionable context (when to use, prerequisites, error recovery). Parameter descriptions are minimal (e.g., 'Optional recall profile' without explaining profile options or defaults). Output schemas are not documented, LLMs cannot predict response structure. No error handling guidance. Tool composition is sound (each does one thing), but descriptions could be 2 - 3x more detailed to match A-grade baselines (50 - 200 chars, LLM-optimized).

• Document output schemas for all tools. For paxm_recall, specify: {hits: [{id, text, provider, score, metadata}], query, total_count}. For paxm_remember, specify: {refs: [{provider, id}], stored_at}.
• Expand parameter descriptions to 50 - 100 chars each. E.g., 'profile: Recall strategy (stm=short-term, ltm=long-term, hybrid=both). Defaults to configured active profile.' instead of 'Optional recall profile.'
• Add context to tool descriptions. E.g., paxm_recall: 'Search memory using the active recall policy. Use for context retrieval before responding. Returns ranked hits with relevance scores.' (currently 50 chars, target 80 - 120).
• Add error recovery guidance. E.g., 'If no hits found, try broadening the query or calling paxm_history to understand available memory.' Include per-tool error classification (retryable, user-fixable, fatal).
• Rename paxm_config_doctor to paxm_check_provider_health or paxm_validate_providers to use clear action verbs and improve LLM intent inference.