who-am-ai

Learns how you want AI assistants to work, from your own messages, and keeps a short profile that your assistants load in every conversation.

Every new session starts blank, so you keep repeating the same things: reply in English, keep it short, do not paraphrase my request, no emojis in commits. who-am-ai reads what you say to your assistants, notes the preferences those messages reveal, and maintains a profile of about 600 words that the assistant loads into every conversation. You say it once; every assistant knows it from then on.

Features

• Learns from what you actually say. It reads only your messages, plus the tail of the reply you reacted to as context, so a short "no, in English" is understood. Assistant text never becomes a preference, and pasted material is ignored.
• A high admission bar. An item enters the profile only when it is general (it changes what an assistant does in unrelated tasks) and durable (a standing rule, or it recurs across chats). Details tied to one ticket or repository stay out.
• You have the last word. Edit me.md anywhere. Your edits are diffed against the last version the tool wrote and become observations with the highest authority, which no inference can override.
• Every change is explained. Each rewrite is stored with what was added, removed or changed and why (whoamai log), and every version of the profile is kept.
• Small and predictable. The profile has a word cap, enforced by code by priority: what you wrote, then what you repeated, then the most recent. There is no daemon, no Docker and no web server.
• Bounded cost. A monthly spend cap stops the LLM calls. A typical user spends a few dollars per month.
• No LLM, no guessing. Judgment is always done by an LLM; there are no heuristic fallbacks. Without a key or at the cap, runs stop and say so.

How it works

Once a day, whoamai run processes the conversations that have settled: it reads your new messages, asks an LLM to annotate the preferences they reveal (each with your literal quote), folds in your edits to me.md, asks an LLM to rewrite the item list (merging duplicates, the latest preference wins, you win over everything) and renders the result with the word cap applied. Details: docs/design.md.

Getting started

Requirements: Python 3.13 or 3.14, uv and an Anthropic API key for the LLM steps. Today the assistant is Claude Code, on macOS, Windows or Linux. On Linux the schedule needs systemd and the key needs a Secret Service (GNOME Keyring or KWallet); without them, run whoamai run yourself.

Install it from PyPI:

uv tool install who-am-ai

whoamai key

whoamai install

key stores the API key in the system keychain (service who-am-ai): the macOS Keychain, the Windows Credential Manager or the Secret Service. install schedules whoamai run once a day at 12:00 and makes Claude Code load the profile by adding @~/.claude/me.md to ~/.claude/CLAUDE.md. To see the first result right away:

whoamai run

To run it from a source checkout instead, use uv sync and prefix the commands with uv run.

Commands

State lives in ~/.who-am-ai/ (override with WHO_AM_AI_HOME).

Development

The tests use a fake LLM and a temporary home, so they need no API key and no network.

uv sync

uv run python -m pytest -q

Documentation

Privacy and security

who-am-ai reads your assistant transcripts (today, Claude Code's) locally and sends your messages, and the tail of the replies they react to, to the Anthropic API to annotate preferences. Nothing else leaves your machine. State is kept in ~/.who-am-ai/, and the API key in the system keychain. See SECURITY.md to report a vulnerability and CONTRIBUTING.md to contribute.

License

MIT