Give your AI agents a memory that lasts.
Orion is a knowledge system that gives any MCP‑compatible agent — Claude, GPT, Cursor, or your own — structured, persistent memory that accumulates expertise instead of starting from zero each session.
Stateless agents waste your time, every time.
Agents forget
Every session starts from zero. No accumulated knowledge, no expertise, no continuity — your last hour of context evaporates the moment the conversation ends.
Context is lost
Decisions, patterns, and relationships between ideas vanish when the window closes. Re-explaining the same architecture, the same constraints, every single morning.
Identity evaporates
Switch models and your agent forgets who it was, what it knew, how you work together. The model is a stateless function — your knowledge shouldn't be.
example · from a real beta galaxy after 2 weeks
A cosmology for knowledge.
Orion organizes everything an agent learns into five nested layers. Galaxy holds everything. The Sun is the agent's persistent identity at its center. Planets are the major knowledge domains. Biomes are the living contexts where work happens. Stardust is the atomic unit — one fact, one decision, one procedure.
The top-level container. One Galaxy per persona, project, or organization. Everything you know lives here.
The persistent identity at the center of a Galaxy. Expertise, continuity score, and working style accumulate here — survives session ends and model switches.
A major domain orbiting the Sun — Engineering, Research, Personal. Each Planet is a distinct context with its own Biomes.
A single working memory — a codebase, a study, a relationship. Where work accumulates day to day.
One fact, decision, or procedure. Typed by cognitive region. Linked into the knowledge graph automatically.
Hot, warm, and cold — by design.
Sub-millisecond lookups. Region-tuned TTLs. The first thing brain.recall checks.
Embedding similarity, RRF-blended with the cache. The "felt sense" of the memory layer.
The canonical store. Every Stardust, every Entity, every edge. Source of truth for the graph.
Built like a memory should be.
Semantic search with RRF
Reciprocal Rank Fusion blends keyword cache, vector similarity, recency decay, and confidence weighting into one ranked result set. Tunable per cognitive region.
Persistent agent identity
Agents survive session ends and model switches. Expertise profiles accumulate. Continuity scoring measures how well a new model picks up where the old one left off.
Multi-agent collaboration
Share a Galaxy across agents and teammates. Parallel workstreams stay in sync through shared Planets and Biomes — RBAC scopes keep each agent's access bounded.
Three calls. One persistent mind.
Orion exposes itself as a Model Context Protocol server. Any MCP client — Claude Desktop, Cursor, your custom agent loop — can orient, think, and recall in three tool calls.
Bind a session to a Galaxy. Returns identity, recent biomes, and warm context.
Write a Stardust record. Auto-extract entities, link the graph, route to a region.
RRF-blended retrieval across all three storage tiers. Returns ranked memories with provenance.