Oort

Open-source memory for AI agents.
Old facts are closed out, not overwritten. Old history moves to cheap storage.
Every recall can explain why it was picked.

---

Why Oort

Agents forget between sessions, and the usual fix is a vector store that only grows. That has two problems. Stale facts keep winning ("lives in Berlin" outranks "moved to Lisbon"), and the bill grows with every message you ever stored.

Oort keeps memory the way a solar system keeps matter:

Tier	Name	What lives there	Recall
Hot	Photosphere	What just happened, and anything pinned	Instant, never searched for
Warm	Kuiper Belt	Short distilled facts from past sessions, fully indexed	A few milliseconds
Cold	Oort Cloud	Older history, zstd-compressed. The text stays, the heavy vector goes.	Comes back when it matters again

When a fact changes, the old one is marked superseded and kept for history; a fact found wrong is marked corrected. Recall returns what is true now, unless you ask for --as-of a date, the --history of a fact, or what was --known-at a time (or just ask "what did we use in March").

Quickstart

Install the CLI (Node 22.15 or newer), then connect your agent. oort setup detects your agents, asks how you want to provide keys, connects them and runs a quick check. Every step can be skipped.

npm install -g oortmem

oort setup also installs the managed copy in ~/.oort/cli (npm i -g --prefix ~/.oort/cli oortmem), the one your agents run, so their hooks and MCP entries keep working when you switch Node versions with nvm, fnm, Volta or asdf. More.

To update later, run oort update: it updates every copy here (the managed copy the Claude Code hooks run included), migrates the store and restarts oort serve.

oort setup
oort recall "where does the user live"
oort explain "where does the user live"   # each stage's rank and score

Every command and flag is in docs/cli.md. To use Oort from your own code (import { createOort } from 'oortmem'): SDK: docs/sdk.md · reference: docs/sdk-reference.md · examples/. To build from source instead: git clone https://github.com/oortmem/oort && cd oort && pnpm install && pnpm build && cd packages/cli && pnpm add -g . (pnpm 12).

Works with

Claude Code (hooks, plus the MCP server), Claude Desktop, Codex, Cursor, Gemini CLI, VS Code, and any MCP client over stdio. oort connect <host> writes the config, backs up the old file and prints the diff. Claude Code guide.

Build with Oort using your coding agent

The oort-sdk skill teaches your coding agent to build Oort into your own apps and agent frameworks: the TypeScript SDK (createOort from oortmem), multi-user apps, the Vercel AI SDK, LangGraph.js and OpenAI Agents JS, serverless, the native Python client (from oortmem import Oort), and the CLI's JSON for other languages. Every code block in it is a tested example. It is a separate plugin, oort-dev, so "help me build with Oort" installs without "memory for my own sessions" (the oort plugin), and the other way round.

Claude Code:

claude plugin marketplace add oortmem/oort
claude plugin install oort-dev@oortmem

(or, in a session, /plugin marketplace add oortmem/oort and /plugin install oort-dev@oortmem)

Codex:

codex plugin marketplace add oortmem/oort
codex plugin add oort-dev@oortmem

(or /plugins in Codex)

Any other agent that reads SKILL.md (Cursor, Gemini CLI, GitHub Copilot, OpenCode and more): copy skills/oort-sdk into its skills folder, for example .agents/skills/oort-sdk in your repository, or run npx skills add oortmem/oort --skill oort-sdk. Details for each tool: skills/README.md.

How it works

flowchart LR
  T[Agent turn] --> S[Scrub secrets]
  S --> W[Writer: distil short facts]
  W --> G{Decide: add, update,<br/>supersede or skip}
  G --> H[(Photosphere<br/>hot)]
  H -- consolidate --> K[(Kuiper Belt<br/>warm)]
  K -- age out --> C[(Oort Cloud<br/>cold, zstd)]
  Q[Recall query] --> F[Lexical + vector<br/>validity filter]
  F --> R[Fuse and rerank,<br/>freshness breaks ties]
  K --> F
  C -- stub and entities --> F
  C -. hydrated when returned .-> A
  R --> A[Memory block<br/>for the prompt]

• Write path. Every turn is scrubbed of secrets first. A small writer model distils it into short facts, and a decision model compares each fact with what is already stored: add it, update it, supersede the old one, or skip it.
• Read path. Keyword (SQLite FTS5) and vector (sqlite-vec) search run together. Facts that are no longer valid are filtered out before ranking. Results are fused, reranked when it helps, and freshness only breaks ties, so a newer fact never beats a better match.
• Local first. Everything lives in ~/.oort on your machine, in one SQLite file plus compressed cold blobs. Nothing is hosted unless you choose Oort Cloud. With no API keys and no Oort Pro login, the local commands make no network call at all (recall falls back to keyword search), and a test runs them with every socket blocked. The one exception is a once-a-day check for a new release, a single request to npm's registry that sends nothing about you; OORT_NO_UPDATE_CHECK=1 turns it off (details).

Packages

Package	What it does
oortmem	The oort command line (setup, connectors, the local MCP server, recall, import and export), and the TypeScript SDK (docs/sdk.md)
@oortmem/core	The engine: secret scrubbing, the write gate, hybrid recall, fusion, tiering and project scope
@oortmem/store-sqlite	Storage on SQLite with FTS5 and sqlite-vec, and zstd cold blobs
@oortmem/writer	Turns conversation turns into short, self-contained facts
@oortmem/decide-jev	Add, update, supersede or skip decisions, using TypeSafe Jev
@oortmem/embed-voyage	Embeddings through Voyage AI
@oortmem/rerank-voyage	Reranking through Voyage AI
@oortmem/local-models	Embeddings and reranking on your CPU (ONNX Runtime): no keys, no account
@oortmem/claude-code	Claude Code hooks and transcript import
@oortmem/mcp	The MCP tools, schemas, instructions and skill, served over any memory backend
oortmem on PyPI	The native Python client (Oort, AsyncOort; docs/python.md) and a launcher for the oort command

Install bundles for Claude Desktop, Claude Code, Codex, Cursor and Gemini CLI are in distribution/.

This repo is the open-source engine and CLI only. The hosted side (oortcloud.dev, Oort Cloud and the Oort Pro relay) is a separate service; the engine and CLI reach it only over HTTPS when you use Oort Pro.

Keys, or Oort Pro

Oort calls three model services: Voyage AI (embeddings and rerank), TypeSafe (decisions) and Anthropic (the writer). Bring your own keys with oort keys set. They stay in the macOS Keychain, or in a 0600 credentials file on Linux and Windows.

Or run with no keys and no account: npm i -g @oortmem/local-models && oort models pull downloads small embedding and rerank models (58 MB, MIT and Apache-2.0) and runs them on your CPU, offline. Saving new memories still needs the decision and writer models. Local models.

Or use Oort Pro: $19/month, VAT included where applicable. No API keys needed. Run oort login, then oort cloud use to keep your memories in Oort Cloud in the EU, with a dashboard, a star map and a connector for claude.ai and ChatGPT. You can export or delete them at any time, or keep them on your machine and use only the relay, which stores no content. Pricing.

Projects and trust

Memories are scoped to a project. By default its id comes from the main repository root, so a git worktree shares memories with its main checkout. A project.id in a repo's oort.config.json is ignored until you run oort project link, because a cloned repo could otherwise use it to read another project's memories. Details in docs/cli.md.

Develop

pnpm install
pnpm verify   # public-repo and copy checks, format, lint, build, typecheck and every test

The export format is documented in docs/export-format.md.

Security

Please report vulnerabilities to [email protected], not in a public issue.

License

Apache 2.0. Made in Gothenburg, Sweden.