A three-layer persistent memory system that gives AI agents durable recall across sessions, machines, and frameworks. Facts accumulate. Context persists. Trust evolves.
Each layer serves a distinct purpose, working together to give agents comprehensive, contextual recall without flooding the prompt.
Always-on identity context. User profile, project state, system preferences — injected every turn without being asked. The agent knows who you are before you speak.
Entity-linked knowledge with trust scoring. Decisions, preferences, relationships — each fact carries confidence that evolves with feedback. Better facts surface higher.
Semantic search over ingested documents and past sessions. Ask a natural-language question, get the most relevant fragments — regardless of when they were stored.
Not a demo. A production memory layer designed for reliability, speed, and the messy reality of multi-session agent work.
Store a fact on your laptop. See it on your server. Shared memory pool across all your agents and machines — no sync, no conflicts.
Facts earn trust through use. Helpful knowledge rises. Outdated or wrong facts decay. The system learns which memories matter.
Standing rules that apply only when relevant. "Always use snake_case" fires in Python sessions, not markdown editing. Context-aware behavior shaping.
Write-ahead log queues operations when the service is unreachable. Drains automatically on reconnection. No data lost, no user intervention.
One API call assembles all layers into a token-budgeted block. No prompt overflow. No manual assembly. Just inject and go.
New facts are checked against existing knowledge. Conflicts surface for human resolution — the system never silently overwrites what it knows.
Sirius Memory speaks the Model Context Protocol — the open standard for tool integration. Any agent framework that supports MCP gets persistent memory in minutes.
All your agents share the same memory pool. A fact stored on Machine A is available on Machine B next turn.