PersonaKit reads the conversations you already have with your assistants, proposes short
memories ("you prefer direct, practical feedback"), keeps the exact quote each one came from,
and serves only what you allow back to your assistants over MCP. Everything runs on your own
machine. There is no PersonaKit server and no account.
Status: pre-release (0.1 in progress). This repository is being built in public.
See docs/ROADMAP.md for what exists and what is planned.
Why PersonaKit
- Every memory has a receipt. Each memory is stored with the quote it came from, in your
own words, so you can see why your assistant believes something about you.
- You have the last word. Approve, restrict, hide or forget any memory. Restricted
memories are never served. Assistants can read and propose, but only you can decide.
- Local by design. One SQLite file on your machine. No account, no PersonaKit server,
no telemetry. Extraction uses a model you already use, and sends it only the turns you wrote.
- Small enough to read. A short pure-Python core with no required dependencies, so you
can check what it does before you trust it with your conversations.
Quickstart
Requires Python 3.10+ and a signed-in claude or codex CLI for extraction.
git clone https://github.com/Yanir-R/personaKit.git
cd personaKit
pip install -e ".[mcp]"
personakit sweep --cli claude --model <model-name> --max-calls 20 # read sessions, extract memories
personakit review # optional: approve, restrict, hide, forget
personakit list # see what is stored
personakit stats # counts only, never memory text
personakit serve # MCP server over stdio for your assistants
personakit forget <memory-id> # delete a memory and its quotes for good
sweep, list and stats also take --json for stable, sorted machine-readable output.
The model name is always yours to choose. Nothing leaves your machine except the turns you wrote,
sent to the model you named. See docs/PRIVACY.md.
Connect your assistant
personakit serve is a standard MCP server over stdio, so any client that can launch a local
MCP server can use it.
Claude Code
claude mcp add personakit -- personakit serve
Codex (~/.codex/config.toml)
[mcp_servers.personakit]
command = "personakit"
args = ["serve"]
Your assistant gets two tools:
| Tool | What it does |
|---|
personakit_read | Returns the memories relevant to a question, each with its evidence quotes and whether you have confirmed it. |
personakit_propose | Lets the assistant suggest a new memory. It is stored as unconfirmed until you review it. |
How it works
- Ingest reads local session files. Only the turns you wrote leave the parsing step.
- Extract asks a model you already use, through your own signed-in
claude or codex CLI,
to propose memories. The CLI runs with its tools off, in an empty temporary directory.
- Store keeps each memory with its source quote, a state, and a scope.
- Review is optional. Memories you never review are served marked unconfirmed; you
can approve, restrict, hide or forget any of them at any time.
- Serve exposes the store to assistants over MCP, with the evidence attached.
A memory is always in one of four states:
| State | Served to assistants? |
|---|
unconfirmed (default) | Yes, labelled as unconfirmed |
approved | Yes |
restricted | Never |
archived | Never |
Limits in 0.1
Stated plainly, so you know what you are installing:
- Extraction is a model's guess. Review is how you correct it.
- Ranking is word overlap, on whole words in Latin letters and digits. Questions and memories
in other scripts are not matched yet; approved memories are still returned as standing context.
- A decision cannot be changed, only undone by forgetting.
review shows unconfirmed memories;
there is no command yet to edit a memory's wording or move a decided memory to another state.
- A memory an assistant already received stays in that conversation. Restricting it applies to
future reads.
- Paste and credential detection is by shape, not by meaning. Look before you share a store.
What works today
| Status |
|---|
| Claude Code sessions | Available in 0.1 |
| Codex sessions | Available in 0.1 |
| Any MCP client (read and propose) | Available in 0.1 |
| ChatGPT, claude.ai and Gemini export import | Planned for 0.3 |
Documentation
Contributing
Issues and small pull requests are welcome. Read CONTRIBUTING.md first;
commits are signed off under the DCO. To report a security problem, see
SECURITY.md.
License
Apache-2.0. See LICENSE. The PersonaKit name and logo are covered by
TRADEMARK.md.
