Persistent memory for AI agents · v0.1

Give your AI agents a memory that lasts.

Orion is a local‑first 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.

Local-onlyMCP nativeSource available · FSL
Sun
agent identity
Planet
Engineering
Planet
Research
Planet
Personal
Planet
Projects
Galaxy · my-brain
The problem

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.

How it works

A cosmology for knowledge.

Orion organizes everything an agent learns into six nested layers — borrowed from the structure of the universe itself. Each layer has its own lifecycle, its own queries, and its own way of decaying.

01
Galaxy
A complete knowledge universe

The top-level container. One Galaxy per persona, project, or organization. Survives forever.

02
Sun
Agent identity & values

A steering document — who the agent is, what it values, how it speaks. Read on every brain.orient.

03
Planet
A knowledge domain

Engineering. Research. Personal. Each Planet groups related Biomes under a shared lifecycle.

04
Biome
A project context

A single living memory — a codebase, a study, a relationship. Where actual work accumulates.

05
Stardust
An atomic record

One fact, decision, or procedure. Typed by cognitive region. Linked into the knowledge graph.

06
Entity
A named thing

People, tools, concepts, decisions, code symbols. Auto-extracted. The graph spine.

POSTGRES · COLDCHROMA · WARMREDIS · HOTbrain
Three-tier storage

Hot, warm, and cold — by design.

RedisHot cache

Sub-millisecond lookups. Region-tuned TTLs. The first thing brain.recall checks.

ChromaDBSemantic vectors

Embedding similarity, RRF-blended with the cache. The "felt sense" of the memory layer.

PostgreSQLRelational spine

The canonical store. Every Stardust, every Entity, every edge. Source of truth for the graph.

Features

Built like a memory should be.

brain.recall

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.

graph.traverse

Knowledge graph

Every record links to named entities and related Stardust automatically. Graph traversal, path-finding, hub detection — all native. The shape of what you know.

7 regions

Cognitive regions

Knowledge is typed by reasoning mode — analytical, procedural, contextual, creative, empathetic, critical, strategic. Each region has tuned cache TTLs and retrieval weights.

sun.identity

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.

docker compose

Local‑first, zero cloud

Postgres, Redis, ChromaDB, Ollama — all in Docker on your machine. Optional Anthropic or OpenAI keys for LLM features. Your knowledge never leaves the device.

19 tools

MCP native

19 tools across brain.*, memory.*, and sun.* namespaces. Claude Desktop, Cursor, custom agents — anything that speaks Model Context Protocol just works.

Cognitive regions
analyticalproceduralcontextualcreativeempatheticcriticalstrategic
Integration · MCP

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.

brain.orient

Bind a session to a Galaxy. Returns identity, recent biomes, and warm context.

brain.think

Write a Stardust record. Auto-extract entities, link the graph, route to a region.

brain.recall

RRF-blended retrieval across all three storage tiers. Returns ranked memories with provenance.

agent.py · python · MCP clientorion · localhost:8787
# 1. Orient — bind this session to a Galaxy
result = await brain.orient(
    agent_name="claude",
    galaxy_id="my-brain",
)
# → Loaded Sun · 4 planets · 1,284 stardust · 312 entities

# 2. Think — store a decision in the analytical region
await brain.think(
    content="We chose PostgreSQL over MySQL for its JSON support"
            " and connection pooling characteristics.",
    region="analytical",
    gravity="planet",
    biome="orion-backend",
)
# → Stardust SD-7c4f · linked: PostgreSQL, MySQL, decision

# 3. Recall — RRF-blended retrieval, three weeks later
memories = await brain.recall(
    query="database architecture decisions",
    limit=5,
)
# → 5 results · 1 from cache · 3 vector · 1 graph hop
Without Orionsession 14 · monday
you
Why did we pick Postgres again?
agent
I don't have context from a previous session. Could you share the requirements and constraints again so I can help you decide?
With Orionsession 14 · monday
you
Why did we pick Postgres again?
agent
Three weeks ago you chose PostgreSQL over MySQL specifically for its JSON support and connection pooling. The decision is in the orion-backend biome — want me to surface the full thread?
recall · 1 hit
Who it's for

Three orbits. One memory layer.

For builders

AI agent developers

Building Claude or GPT-based agents that need to remember user context, past decisions, and accumulated domain knowledge across runs — without re-engineering memory yourself.

Drop-in MCP server
Per-agent identity
Provenance on every recall
For teams

Teams with AI workflows

Multiple agents collaborating on shared knowledge — Planets per team domain, RBAC scopes, Galaxy merge for org-wide synthesis. The shared brain your humans always wanted.

Multi-agent Galaxies
RBAC + audit trail
Cross-galaxy merge
For researchers

Power users & researchers

Running local AI stacks — Ollama, MLX, vLLM — who want structured, searchable, persistent memory without sending data to anyone. Your knowledge graph is yours.

100% on-device
Source available · FSL
Hackable schemas
From the team

Thinking out loud.

All posts →
Product8 min

Introducing Orion v0.1 — persistent memory for AI agents

AI agents forget everything between sessions. Orion fixes that. It's a local-first memory system that gives any MCP-compatible agent structured, persistent knowledge — and it gets smarter every session.

AW
Andy Wu
2026-04-28 · 8 min
Engineering12 min

How Reciprocal Rank Fusion makes agent memory retrieval work

No single retrieval signal — keyword, vector, or graph — is sufficient for agent memory. We use Reciprocal Rank Fusion to blend all three, and the results surprised us.

AW
Andy Wu
2026-04-22 · 12 min
Company5 min

Local-first is a principle, not a feature

Your knowledge graph is a map of how you think. That data should never leave your machine. Here's the architectural, ethical, and practical case for local-first AI memory.

OT
Orion Team
2026-04-15 · 5 min
Observability

Galaxy strength is a number you can watch.

Five dimensions, scored continuously: how much you've stored, how connected the graph is, how fresh the records are, how varied the cognitive regions, how well your Biomes cover your Planets.

Volume78
Density64
Health91
Diversity71
Coverage56
Galaxy strength
72+8
Healthy
VolumeDensityHealthDiversityCoverage

Start building agents that remember.

Five-minute Docker quickstart. Five-step onboarding. One persistent Galaxy that your agents will keep filling, session after session.

Get started — it's free→★ View on GitHub

Runs entirely on your machine. No accounts. No cloud. No data leaves your device.