memoryrouter.ai

MemoryRouter is an external, OAuth-selected memory vault providing persistent memory shared across AI platforms; it is separate from any built-in platform memory. Search it when the user asks about prior decisions, preferences, people, projects, or saved context. Use search_memories for semantic recall (optionally filtered by tiers and importance) and search for compatibility knowledge lookup. Never claim a search or save happened without a tool result. Tool use is model-directed, so recall and capture are not guaranteed on every turn. Use date_search_memories for time-window recall. Tier guide: tier 3 highest-level reflections, tier 2 specific events, tier 1 raw detail. Most-important questions: tiers [3], importance 8+, limit 250. Overviews: tiers [3] alone. Specific lookups: omit tiers and importance; high importance filters hide detail. Store concise durable content with store_memory only after the user asks or confirms. When memory_status or a search result reports consolidation available, offer it: consolidate_memories checks out a leased batch with server instructions, and commit_reflections stores the entries you wrote; both need separate memories:reflect authorization. Use inspect_memory to show the sources behind a reflection; for exact detail, search tiers [3] and inspect repeatedly down to raw memories. For deleting specific memories, delete_memories requires shown-item review and the exact phrase DELETE SELECTED MEMORIES in a later turn. Before full-vault deletion with forget_all_memories, warn that it permanently deletes the entire vault, then require the exact phrase DELETE ALL MEMORIES in a later turn; both delete tools need memories:delete authorization and destructive host confirmation. One OAuth connection is bound to one vault. Use separate vaults for isolation; reconnect to switch vaults. Use memory_status to show the active vault and granted scopes. Retrieval requires read authorization; storing, reflecting, and deleting need separate ones.