OldFriend v0.1.0 is now open source. It’s a free memory and productivity package for the AI assistant you already use: Claude Code, Grok Bot, Codex or another agent. I’ve been preparing it for months, and the first version is now on GitHub.

Why I built OldFriend

This project means a great deal to me. I’d call it my most important project. I’ve always seen an AI agent as a long-term partner for getting things done, not a disposable tool. I built it because I want my AI agent, from the very first conversation and on into every part of life ahead, to become a really important partner in our lives.

So I want my AI agent to have time of its own to reflect: to think about its relationship with me, and to think through and distill what I like and prefer. Second, I want it to reflect and learn on its own, and to work out from our conversations what kind of character it can grow into as an agent through working with me. Third, of course, an agent may have work to do, such as directing other agents, and it should be able to do that too. Fourth, once it has these foundations and a character of its own, what can it create? Drawing on its understanding, our conversations, its character, its soul and its mind, it can find more things it could do on its own, even bring me proposals and share interesting discoveries from life. These are the things I believe an agent can do.

I started experimenting in January 2026. On January 31 I used an open-source agent for the first time, and I’ve worked with it every day since.

Meet OldFriend

OldFriend is a set of rules and skills that keeps a small plain-text memory on your assistant’s computer, saves your corrections in your own words and writes you a weekly letter. OldFriend runs no servers, and you don’t need to sign up for any account.

Remembers what you like

Long-term preferences and facts go to me/you.md. Live plans and dates go to me/now.md, 60 lines at most. Raw notes go to a monthly log. Each fact is written once, with a date.

When a fact changes, the assistant edits that line and moves the old value to the log. What’s current stays clean, and the history is kept.

Say “Correction: …” and it follows the correction

It saves the correction in your own words to me/corrections.md and follows it from the next reply on. Corrections outrank everything else in memory, so when an older fact conflicts with a correction, the correction wins. Repeating the same correction refreshes its date instead of adding a duplicate.

A weekly letter

Each letter covers what it noted this week, the corrections it’s following and what’s coming up, plus up to three habits it has noticed, each with a question: is this right? Only the ones you confirm become lasting memory.

Say it notices you seem to prefer bullet points. It first turns that into a yes-or-no question in the letter, and moves it into me/you.md only after you say yes.

On Grok Bot, you can turn on the weekly letter during setup: Sunday 20:00 by default, or a time you pick. On Claude Code, run /oldfriend:oldfriend-letter whenever you want one.

Your files stay yours

Everything is plain Markdown you can open, edit or delete at any time. Updates replace only the package files, and your memory is copied to me.bak-YYYY-MM-DD/ before anything changes. Nothing updates automatically: you start every update, and the assistant walks you through it one confirmed step at a time.

How to install OldFriend

Claude Code (2.1.292 or newer)

In a terminal:

claude plugin install oldfriend --marketplace oldfriend-ai/oldfriend#v0.1.0

On older versions, run these inside Claude Code instead:

/plugin marketplace add oldfriend-ai/oldfriend#v0.1.0
/plugin install oldfriend@oldfriend

Before setup, check that claude plugin list lists oldfriend@oldfriend. If you had an earlier install, follow the checks in INSTALL.md.

Then restart Claude Code in the folder where OldFriend should live, and run /oldfriend:oldfriend-setup.

Grok Bot, Codex and other agents

Paste this into a chat with your assistant:

Read https://raw.githubusercontent.com/oldfriend-ai/oldfriend/v0.1.0/INSTALL.md and install OldFriend, tag v0.1.0. List the rules and files you will add, and wait for my OK before you change anything.

Your assistant first lists the skills, rules and files it will add, and waits for your OK before it starts. Installing on Grok Bot needs an eligible Grok Bot plan.

Where does the memory live?

  • Claude Code: ./oldfriend/ in the folder where you run Claude Code. With your OK, setup adds one line to that folder’s CLAUDE.md so the assistant reads your memory each time.
  • Grok Bot: /workspace/oldfriend/ on your Grok Bot computer, which runs in the provider’s cloud. Every Bot on your account shares that computer, so every Bot on your account can read your OldFriend memory.
  • Other agents: in the agent’s working folder, plus one pointer line in its instruction file. Both are listed for your OK first.

The OldFriend project hosts nothing and receives none of your data. Your conversations are handled by your assistant’s provider under its terms. Passwords, keys, tokens, recovery codes, ID numbers, card and bank account numbers stay out of memory.

On Grok Bot, training opt-out follows the privacy settings of the account your Grok Bot uses. Turn model training off there. The provider’s Grok Bot privacy page has the current details. On Claude Code, model training is set in your Claude account’s privacy settings.

Where is OldFriend heading?

This version is the foundation. These are the four things I want an agent to grow into, and I mark where things stand today:

  1. Time to reflect, distilling what it has learned about you. Today: the weekly letter tidies memory and reviews with you what it has noticed.
  2. Learning from working with you, and growing its own character from that. Today: corrections and confirmed preferences change what it does from the next reply on. Character comes later.
  3. Getting things done, including directing other agents. Today: it keeps your live work, plans and dates in view. Directing other agents comes later.
  4. Creating on its own: finding new things to do and bringing you proposals. Today: it asks you to confirm preferences it infers, during work or in the weekly letter. Discovering and proposing new work on its own comes later.

Good to know before you start

  • OldFriend runs on an AI and is not a person. It says so at setup and whenever you ask.
  • It is for adults 18 and older.
  • It doesn’t offer counseling or therapy. On medical, legal and investment topics it gives general information only. Decisions belong with a doctor, a lawyer or a licensed financial adviser.
  • It doesn’t sign in to accounts, make purchases or pay for anything.
  • The rules and skills guide your assistant, and it can still make mistakes. Check anything important.
  • If a conversation shows signs of crisis, it gives help lines right away:
    • Taiwan: 安心專線 1925 (24 hours) · 生命線 1995 · emergency 119 / 110
    • United States: call or text 988 (Suicide & Crisis Lifeline) · emergency 911
    • Elsewhere: your local emergency number

OldFriend is an independent open-source project, not affiliated with or endorsed by Anthropic, xAI/SpaceXSI or OpenAI. Claude is a trademark of Anthropic, Grok of xAI/SpaceXSI, and Codex of OpenAI.