Three tools with clear, actionable descriptions (avg 85 chars, within 10-1024 baseline). All have input schemas with type definitions. However, output schemas are undocumented, LLMs cannot predict return structure. Parameter descriptions are minimal (e.g., 'The handle or keyword query' lacks format/constraint guidance). No error handling guidance, no recovery paths. Naming is verb-forward (archive, recall, status) but generic, 'lethe_*' prefix adds namespace clarity. Missing idempotency hints, rate-limit guidance, and per-parameter validation rules.

• Document output schemas explicitly. For lethe_archive, specify: returns {handle: str, tokens_saved: int, label: str|null}. For lethe_recall, return {content: str, label: str, archived_at: str} or error {error: str, suggestion: str}. For lethe_status, return {total_archived: int, total_tokens_offloaded: int, items: int}.
• Add format/constraint guidance to parameters. E.g., 'query_or_handle: A short alphanumeric handle (e.g., "abc123") or a keyword phrase (1-50 chars). Handles are case-insensitive; keywords are substring-matched.'
• Implement structured error responses. Instead of returning '[lethe: nothing found]' as a string, return {error: 'not_found', message: 'No archive found for query. Try a broader keyword or check lethe_status() to see available archives.'}.
• Add idempotency and retry guidance to lethe_archive description: 'Idempotent: calling with identical content and label returns the same handle. Safe to retry on network failure.'
• Enhance tool descriptions with composition hints. E.g., lethe_archive: 'After calling a verbose tool, use this to offload output. Then pass the returned handle to lethe_recall when needed.' lethe_recall: 'Retrieve content archived by lethe_archive using its handle or keyword search.'