Three tools with visible schemas and descriptions, but significant quality gaps. Tool names lack clear action verbs (memory_add vs add_memory). Descriptions are lengthy (200+ chars) and written in Chinese with domain jargon, making them harder for LLMs to parse reliably. Parameters lack type constraints (enums for 'kind', 'migration', 'scope'). No output schema documentation. Error handling is minimal, tools return free-text strings rather than structured responses. No parameter validation guidance. The 'kind' parameter accepts 'fact/procedure/handover' but this is buried in description text, not formalized as an enum. Overall, definitions are present but underspecified for production agent use.

• Rename tools to verb_noun: add_memory, search_memory, list_preload_candidates. This matches LLM conventions and clarifies intent.
• Rewrite descriptions in English, 50 - 150 chars, following WHAT/WHEN/WHY pattern. E.g., 'Add a memory to the cross-device store. Use for facts, procedures, or handoff notes. Tags and migration strategy are optional.'
• Add enum constraints to schema for 'kind' (fact|procedure|handover), 'migration' (local|edge|cloud|auto), and 'scope' (tag:*|kind:*). Remove free-form string types.
• Document output schema as JSON with typed fields. E.g., memory_add returns {node_id: string, scene: string, migration: string, kind: string, message: string}. memory_search returns {results: [{node_id, content, score}], total: int, injected_context?: string}.
• Add error classification: return {error: string, code: 'not_found'|'invalid_param'|'internal', recovery: string}. E.g., 'No memories found. Try a broader query or check scope filters.' This guides LLM retry logic.

Inferred effective spec: <=2025-11-25.

• Relies on Logging (deprecated) - log to stderr or use OpenTelemetry