Tallyfold

Tell your agent once.

Tallyfold gives coding agents memory that lasts across sessions, projects and machines. It lives in a private git repository you own, as plain Markdown you can read and edit.

Install for Claude Code Read the source

A split tally stick Two halves of a hazel stick with matching notches. The halves part, a new notch is cut through both at once, and they close again. on your machine in your repo
The English Exchequer kept accounts on hazel tally sticks. A stick was notched, then split along its length, and each party kept one half. The notches matched when the halves were put back together. Tallyfold keeps memory the same way: each fact your agent learns is cut into the copy on your machine and the copy in your repository at once.

How a fact travels

  1. You tell your agent something worth keeping

    A correction to how it works, a preference, who owns a service, why a decision went the way it did.

  2. It writes one line to your memory

    Tallyfold adds the line to a Markdown file, stamps the date and the client that wrote it, commits it and pushes it to your repository.

    MEMORY.md
# Memory

- Prefers a table when comparing options, with the recommended one first [added: 2026-10-07; by: claude-code]
- Never polls CI in the foreground; reports the run URL and stops [added: 2026-10-07; by: claude-code]

## Index

- [[people]]
- [[projects/payments]]

  3. The next session already knows

    When a session starts, Tallyfold pulls the latest copy and puts MEMORY.md in front of the agent. That works in any project and on any machine where the plugin is installed. Topic files load when the agent needs them.

  4. When the fact changes, the line changes

    The agent replaces or removes the old entry instead of adding a contradiction. Git keeps every earlier version, so nothing is lost and any change can be undone.

What your agent keeps

Saved

  • Corrections to how it works
  • Your preferences in writing, tools and code
  • People, teams and who owns what
  • Facts about your projects that outlast one task
  • Decisions and the reasons behind them

Skipped

  • Anything the code or docs already say
  • Details that only matter for the task at hand
  • Guesses it hasn't confirmed
  • Passwords, tokens and keys, which Tallyfold refuses to write

Plain files you own

Your repository
Memory lives in a private git repo on your own account. Tallyfold runs on your machine and has no server of its own.
Readable and editable
Each fact is one Markdown bullet. Read it on GitHub, fix a line by hand, or delete a file you no longer want.
Every change is a commit
You can see when a fact arrived, which client wrote it, and what it said before it changed.
Small at startup
MEMORY.md stays under 4,000 characters, so sessions start with what matters and look up the rest.
Open formats
Files follow the Agent Memory Repo format. The plugin follows the Agent Plugins 1.0 specification, so any agent that reads those formats can use your memory.
Safe with concurrent sessions
Two sessions writing at once both land. If they change the same line, the later write is undone and the agent is shown both versions to merge.

Install

Tallyfold runs on macOS and Linux and needs git and curl. Claude Code is the tested client.

1. Add the plugin to Claude Code

claude plugin marketplace add swiftugandan/tallyfold
claude plugin install tallyfold@tallyfold

2. Make an empty private repository for your memory

gh repo create agent-memory --private

3. Ask your agent to set it up

Set up Tallyfold with git@github.com:you/agent-memory.git

Tallyfold reaches the repository with your machine's own git credentials. To try it without a remote, ask your agent to set up Tallyfold locally; you can add a repository later.

Other agent clients

The plugin folder is an Agent Plugins 1.0 package, which VS Code, Cursor, GitHub Copilot and Codex can install from the repository. These clients haven't been tested yet. Their agents load memory through the Tallyfold skill when a task starts, because the specification has no session-start hook.

Next: memory you share

This release keeps personal memory. The same design extends to two shared kinds, and the code already has the places they plug in.

Team memory
A repository your team reads, where your agents propose facts on a branch and they reach everyone after checks and review.
Swarm memory
A branch shared by agents working one problem in parallel, so each builds on what the others have found.