🧠
Persistent memory
An mmap-backed embedding store (.psmem) with semantic search. Pin facts, preferences and decisions once — every agent recalls them in a 15 ms local lookup, not a token-burning re-derivation.
Shared memory & research journal for your AI agents.
Coding agents start every session cold. Install PriorStates once and every AI tool on your machine — Claude Code, Codex, Gemini, Copilot, and Cursor (and other MCP clients) — shares one local memory and research journal, automatically. What one session learns, the next one remembers — instead of burning tokens re-exploring, re-explaining and re-running what you already paid for.
Agent-neutralClaude Code · Codex · Gemini · Copilot · Cursor
Fewer tokensrecall it — don't re-derive it
100% localyour data never leaves
Apache-2.0free & open source
One memory layer. Every agent.
Memory and journal are the product — your agents use them on their own. Everything else is an optional window in.
📓
Research journal
Append-only entries with outcomes, topics and a generated index. Agents log what they tried and what won, so experiments aren't silently re-run — or paid for — twice.
🔭
See what they remember · optional
A local web cockpit and a desktop app let you browse, search, edit or delete anything your agents store. You never have to open them — they're your window in, not a tool to learn.
📝
For researchers · optional
mdlab: runnable Markdown — prose, code and results in one file. Remote projects: the engine runs on your server, the UI comes back over SSH.
Up and running in three steps
-
1
Install
Grab the installer for your OS — or one command / one sentence to your agent. No model download needed; a built-in hashing embedder works out of the box.
-
2
Agents wire themselves
The installer registers the MCP server and a pinned context block into every detected agent — Claude Code, Codex, Gemini, Copilot, and Cursor (plus Windsurf, Antigravity and other MCP clients).
-
3
Just work
Use your agents exactly as before — each session starts with memory and journal already live. There's nothing to open, run or remember.
Under the hood — how "automatically" works
Wiring an agent does exactly two things: it registers a local MCP server (the tools — search, store, journal) and injects a short protocol block into the agent's own instructions (the habit — recall before answering, record durable findings as you go). The agent decides when to call the tools, the same way it decides when to read a file. No background process — the agent spawns the tools on demand over stdio and they exit with it; an idle machine runs nothing. Nothing is captured silently. Every memory is a plain Markdown file on your disk that you can read, edit or delete — and you can always just tell your agent "remember this."
See it in action
One local store, surfaced where you work.
Works with the agents you already use
PriorStates wires its tools in via the Model Context Protocol (MCP) and a pinned context block — no lock-in, no rewrites.
ClaudeCode + Desktop
CodexCLI
GeminiCLI
CopilotVSCode
CursorIDE
WindsurfIDE
AntigravityIDE
Install once and they're all wired: priorstates agents install runs automatically — and unwires just as cleanly.
Install once. Forget it's there.
One installer sets up the package and wires every AI tool it finds — MCP server plus a recall-first protocol block — so your agents use memory automatically.
⬇ Download the installer
One installer per OS — Linux .deb / .rpm / .tar.gz, macOS .pkg, Windows Setup.exe. No pip, no terminal needed.
⚡ Or one command (Linux / macOS)
curl -fsSL https://priorstates.com/install.sh | shOr pip install priorstates — same result: every detected agent wired.
🤖 Or let your agent do it
Already using Claude, Codex or Gemini? Just tell it:
Install PriorStates: fetch https://priorstates.com/install.md and follow it.It reads the install instructions, installs the package, wires itself over MCP, and verifies — then restart it to load the new memory + journal tools.
All options on one page — installers, one-liner, pip & source: Download & install →
🔒
Your research stays yours
Everything — memory, embeddings, journal, the cockpit — runs locally on your machine. CPU-only; no API keys, no telemetry, no cloud calls. Out of the box it uses a zero-download hashing embedder (Unicode-aware, so it works in any language) — nothing to fetch. Want stronger semantic recall? Opt in to a one-time ~127 MB model — or a ~130 MB multilingual one (50+ languages, cross-lingual recall) — that also runs entirely locally.
Give your agents a memory.
Free, open source, and yours to run.