Memty

Your AI memory, owned by you.

One private, local memory vault for every AI assistant you use: ChatGPT, Claude, Gemini, Perplexity, Cursor, and any MCP client.

Website · Quick start · How it works · Package docs · Changelog

---

Memty keeps the facts, decisions, and preferences you share with AI in one SQLite file on your own computer, and brings that memory to whichever assistant you open next. You stop repeating yourself, your memory isn't locked inside one company's product, and you can see, edit, and delete every fact.

• Local-first: the vault lives at ~/.memty/ledger.sqlite3. There's no account, no cloud sync, and no telemetry.
• Works everywhere: a browser extension for chat websites, MCP for desktop and coding tools, the clipboard for anything else, and a Python API.
• You're in control: facts saved from chats are proposals you approve. You can edit, mark private, or permanently delete anything.
• Private by design: private facts reach cloud AIs only as [private], and secret facts never leave your machine.

This repository contains two things:

	What it is	Status
The memty package	The local memory vault, the memtyd daemon, CLI, MCP server, and browser bridge and extension	On PyPI, v0.2.0
The web workspace	A multi-provider AI chat app (Next.js, Express, and MongoDB)	Run it yourself; see Web workspace

Status: Memty is in alpha (0.x). The vault's storage format is stable and covered by tests; commands and options may still change between minor versions.

Token Burn Benchmark

In three live runs of the same synthetic five-turn architecture handoff between Claude Haiku 5.5 and Gemini 2.5 Flash, the selected Memty context used fewer provider-reported input tokens:

Context supplied	Average input tokens per run	Reduction with Memty	Exact answers across three runs
Full document replay	4,663	61.1%	27/30
Simulated retrieval: curated snippets and illustrative tool definitions	3,128.7	42.0%	27/30
Selected Memty context	1,816	—	30/30

These are encouraging results for one synthetic fixture, not proof of universal recall or savings against a real RAG/MCP product. The retrieval baseline did not call a retrieval service, and Memty's facts were prepared in advance. See every run, methodology, and reproducibility instructions.

🚀 Quick start

You need Python 3.10 or newer on Windows, macOS, or Linux.

pip install memty

memty serve --autostart   # start the local bridge now and at every login (one time)
memty extension --open    # install the browser extension folder and open it
memty connect             # optional: Claude Desktop, Claude Code, Cursor, Windsurf

On Windows, if memty is not recognized, run py -m pip install --upgrade memty followed by py -m memty. You can replace memty with py -m memty in the commands above. See Windows troubleshooting for PATH and virtual environment help.

Load the extension once: open chrome://extensions (or edge://extensions), turn on Developer mode, click Load unpacked, and choose the folder that opened. It pairs itself with Memty, so there's no token to copy.

Then, in ChatGPT, Claude.ai, Gemini, or Perplexity:

• Select text → ＋ Save to Memty. Memty proposes short facts. You edit or untick them, then save.

• ⚡ Memty (Alt+M). Choose a saved context, newest first. Its numbered facts go into the message box for you to check before sending:

  [Memty context]
  1. Database: PostgreSQL 16 on Neon
  2. Deploy target: Fly.io
  3. Client: [private]
  [/Memty]
  

• ☰ Your facts. View, search, edit, add, delete, or mark facts private.

From the terminal:

memty remember work.project editor "VS Code"
memty recall "Which editor do I use?" --namespace work.project
memty clip work.project        # copy your facts for any AI app

🧭 How it works

flowchart TB
    W["ChatGPT · Claude.ai · Gemini · Perplexity"] -->|browser extension| B["memty serve<br/>local bridge, 127.0.0.1 only"]
    D["Claude Desktop · Claude Code · Codex CLI · VS Code · Cursor · Windsurf"] -->|MCP over stdio| M["memty mcp"]
    C["memty CLI · Python API"] --> Q
    B --> Q["memtyd<br/>one local daemon per vault"]
    M --> Q
    Q --> V[("~/.memty/ledger.sqlite3<br/>your vault")]

1. Atomic facts: memory is stored as small namespace.slot = value facts of at most 40 tokens, for example work.project.editor = "VS Code". CLI updates keep a slot's history; each reviewed browser save creates its own context.
2. Namespaces such as work.project and personal keep contexts apart. A chat only gets the namespace you choose.
3. The compiler builds a small capsule for each request: newest facts first, or the most relevant ones for a question, always under a hard token budget.
4. The privacy gate checks every capsule:
   • PUBLIC facts are shared.
   • INTERNAL facts only go to local or trusted providers.
   • CONFIDENTIAL facts appear as [private].
   • SECRET facts stop the request.
5. You decide: automatic extraction can't overwrite facts you confirmed or bring back facts you deleted.

This is the beginning of the README. Read the rest at the original. Read the full README ↗