Install PersonaKit, run a first sweep, review memories, and connect an assistant over MCP.

Start with PersonaKit

Sweep your sessions. Look at what it proposed. Connect an assistant.

Requirements

• Python 3.10 or newer.
• A signed-in claude or codex command-line tool, used for extraction.
• You choose the model name. PersonaKit never defaults one.

1. Install

git clone https://github.com/Yanir-R/personaKit.git
cd personaKit
pip install -e ".[mcp]"

The mcp extra is only needed for personakit serve.

2. Sweep

personakit sweep --cli claude --model <model-name> --max-calls 20

--max-calls caps model calls for the run (default 50). Session files that have not changed since the last sweep are skipped. Sessions with fewer than three of your own turns are too thin and produce nothing. Only turns you wrote are sent to the model. See Trust and data boundaries.

3. Look at what was stored

personakit list
personakit stats

stats prints counts only and never memory text. list --state <state> filters by state, and sweep, list and stats take --json for machine-readable output.

4. Review (optional)

personakit review

See Memory controls. Skipping review is fine: memories are served labelled unconfirmed.

5. Connect an assistant

personakit serve

This runs an MCP server over stdio. Point your assistant's MCP configuration at that command:

claude mcp add personakit -- personakit serve

For Codex and other clients see Connections.

6. Check the loop

Restrict or hide one memory, ask the assistant a related question, and confirm it no longer appears. Restricted memories are counted as withheld and never shown.

Remove everything

personakit forget <memory-id> deletes a memory and its quotes. Deleting the store file removes all data.