MemoSync GitHub

for Claude Code, Codex, Pi & OpenCode

Co-manage memory with your coding agent across sessions.

Download for Mac See how it works

For Apple silicon Macs · or run it from source

mentioned in Awesome Agents Awesome Agent Memory Awesome Context Engineering Awesome Agent Ecosystem

MemoSync in the middle of a turn: proposed memory changes, the working memory for the turn, a reply citing its memories, the memory use audit, and the turn's memory record.

context — this session

memory — stays

Session 1Mon 09:12

Set up the checkout service. This repo uses bun — never npm.

On it. Installing dependencies with bun.

bun add stripe hono

bun test 12 passed

Checkout service is scaffolded under services/checkout.

Session 2Thu 15:40

Add Stripe checkout next to the PayPal flow.

Working memory for this turn M-02 M-07

Installing stripe with bun M-02

Totals stay in integer cents M-07

npm run buildstopped

bun run build M-02

Memory use audit

M-02violated, then corrected

M-07shaped this turn

M-11not applicable

M-07Money stays in integer cents.project
M-11Commit messages in English.personal
M-02Use bun, never npm.proposed
M-19Run tests with bun test.auto · revert
M-20Deploy keys live in 1Password.sensitive · waits

you

your agent

Automatic

01 / 07 · Learn

It learns while you work.

MemoSync notices what is worth remembering in the conversation and proposes it as memory.

  1. Learn: MemoSync notices what is worth remembering in the conversation and proposes it as memory.
  2. Review: nothing is kept until you say so. Every proposed change waits for review.
  3. Persist: context clears, memory stays. The memory board carries what matters into the next session.
  4. Choose: before a turn starts, you confirm which memories go in and how they will be used.
  5. Follow: replies cite the memories they apply, and you can stop the agent when it drifts.
  6. Check: a separate audit marks each memory as shaped, violated, not applicable or no visible effect.
  7. Automatic: safe changes apply on their own; sensitive ones wait; the audit speaks up only when something broke.

the full story, on film

Watch one turn, start to finish.

bring your own agent

Runs the agents you already use.

Use your own sign-in, or an API connection to any provider. Engines on an API connection run in a separate profile, so your own ~/.claude and ~/.codex are never touched.

Claude Code

Your installed claude, or the copy bundled with the Agent SDK.

Your sign-in · Anthropic API · OpenAI Chat via translation

Codex

Your installed codex, driven through codex app-server.

Your sign-in · OpenAI Responses · OpenAI Chat via translation

Pi

Built in when you run from source. The Mac app uses your installed pi.

OpenAI Chat · OpenAI Responses

OpenCode

Your installed opencode, version 1.17.4 or newer.

Your sign-in · Anthropic API · OpenAI Responses · OpenAI Chat

…and more than seventy presets, with endpoints and models filled in.

two ways in

Download it, or run from source.

Download for Mac

Apple silicon. Not notarized yet: the first time, open it from System Settings → Privacy & Security → Open Anyway.

  1. Have an agent ready: Claude Code, Codex, Pi or OpenCode, with its own sign-in or an API key from any provider.
  2. Connect it in Settings → Models.
  3. Open a project and send a task. Review what it proposes, and the memory builds up as you work.
from source
git clone https://github.com/SLEEPYBQ/MemoSync.git
cd MemoSync
bun install
bun run build
bun run start

macOS or Linux, with Bun 1.3.5 or newer. Opens at localhost:3210.