Bloom provides well-structured tool definitions with complete JSON schemas, clear descriptions, and thoughtful parameter constraints. All 6 tools have descriptions (avg ~100 chars), input schemas with types, and reasonable defaults. Naming follows verb_noun convention (recall, remember, recent, forget, sessions, stats). However, descriptions lack actionable context about when to use each tool vs. alternatives, and error handling guidance is minimal. Output schemas are documented in code but not exposed to clients. No tool annotations (readOnlyHint/destructiveHint) despite clear risk profiles.

• Add tool annotations to MCP tool definitions: mark 'recall', 'recent', 'sessions', 'stats' as readOnlyHint=true; mark 'remember' as destructiveHint=false (write but not destructive); mark 'forget' as destructiveHint=true (reversible but destructive intent).
• Expand tool descriptions to include disambiguation: 'Use recall() for semantic search across all sessions; use recent() to fetch the last N turns from a specific session in order; use sessions() to list available sessions before filtering.'
• Document output schemas in MCP tool definitions. Add outputSchema field with properties: {ok: boolean, count?: integer, results?: array, id?: integer, ts?: integer, error?: string}.
• Add error recovery hints to remember() description: 'If content is too large, split into multiple calls or compress. If embedding fails, the turn is still stored without a vector.'
• Rename 'stats' to 'get_database_stats' and expand description: 'Returns database statistics: total turn count, session count, database size in bytes, and schema version. Use to monitor memory usage and plan archival.'
• Add parameter relationship documentation: 'session_bias and filter_session are mutually exclusive in recall(). Use session_bias to boost results from a session; use filter_session to restrict to only that session.'

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

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