MemoSync

Co-manage your coding agent's memory across sessions — see what it learned, what it used, and whether it followed it.

Stars
★ 396
Last updated
6d ago
License
MIT
Primary language
TypeScript

At a glance

How it runs
Desktop appWeb appCLI
Works with
Universal · cross-platformCodex · Claude Code · OpenAI API · Claude API
Cost
Free software; you pay for model usage
Setup effort
Medium · a few setup steps
You'll need
Bun 1.3.5+Claude Code or Codex installed and signed in, or an API keymacOS on Apple silicon for the DMG build@earendil-works/pi-coding-agent 1.3.0+ for Piopencode 1.17.4+ for OpenCodeShell / CLINetwork accessLocal filesystem
Typical use
Developers already running Claude Code or Codex who cannot tell what has accumulated in CLAUDE.md or which items shaped this turn.
Not a fit if
  • Users who want zero-review, fully automatic memory handling
  • Teams needing a native Windows or Intel Mac desktop build
  • Teams that cannot send conversations to a model provider
Website
memosync.cc ↗ set on GitHub
Maintenance
Commits in the last 30 dayslast commit 2026-11-15latest release v0.45.0 (2026-11-14)+14.7 ★/day over 7 days

What does this agent do, and when should you use it?

MemoSync is a local workbench that surfaces the memory Claude Code, Codex, Pi and OpenCode already keep — CLAUDE.md files, extracted notes, past lessons — as a reviewable library. Memory is stored as short items with a personal, project or session scope, a version and a history, and a single turn walks those items through four stages: reviewing proposed changes, choosing working memory, citing memories inside replies, and an after-the-fact audit. It ships as a React client plus a Bun server that opens at http://localhost:3210, and can also be packaged as a macOS desktop app. Everything lives under ~/.memosync/data/ on your machine as SQLite plus a Markdown copy, and MemoSync uploads nothing itself, though the conversation and selected memories still go to your model provider.

Before a turn, MemoSync proposes new memory candidates drawn from the last exchange and from what the agent actually did (files, commands, errors), lists transfers from other projects as three side-by-side versions (original rule, generalized rewrite, project adaptation), and surfaces fixes to existing items such as conflicts, duplicate merges and stale or repeatedly broken rules; nothing reaches the store until you accept, edit, rescope or dismiss it. A Working Memory for This Turn card then lists the items the agent intends to use with a one-line reason each; you can remove rows, add others from the library, or ask in plain words, and the agent starts only after you confirm. During execution replies cite memories as M-07, tool calls fold into lines like Read 3 files, ran 2 commands, and pressing Stop on a citation lets you write the correction and optionally enforce the item for the rest of the run, with the agent resuming from the interrupted sentence. When the turn ends, a separate audit pass reads the whole exchange and rates each selected item violated, shaped, not applicable or no visible effect, tagging whether the verdict came from the audit or the agent itself, with Where used jumping to the backing sentence and tool call. Between turns the Memory Board shows the whole library as a graph or as one column per scope, where items can be dragged between scopes, edited, read as history or archived; Memory files keeps a Markdown copy and can import an existing CLAUDE.md as candidates, and Import memory splits what Claude Code and Codex already remember into items for review. The Memory Record logs every turn's proposals, selections, citations, stops and verdicts and survives context compaction.

  1. Developers already running Claude Code or Codex who cannot tell what has accumulated in CLAUDE.md or which items shaped this turn.
  2. Teams that want every memory change to pass a human review before it edits project rules.
  3. People working across several repositories who want a general rule transferred from one project to another, compared version by version.
  4. Context-engineering practitioners who need to measure whether a given memory shaped a turn, had no effect, or was violated.
  5. Solo developers who want memories and chat history to stay on their own machine.
  6. Anyone who needs to interrupt an agent mid-task when it contradicts a memory, and keep that correction attached to the item.

How do you install or deploy this agent?

On macOS (Apple silicon) download the DMG from Releases and drag MemoSync to Applications; macOS blocks the first launch until you choose Open Anyway in System Settings → Privacy & Security. The app bundles no agents, so install Claude Code or Codex first, and install @earendil-works/pi-coding-agent 1.3.0 or newer to use Pi.

From source you need Bun 1.3.5 or newer plus a signed-in Claude Code or Codex, or an API key from any provider:

git clone https://github.com/SLEEPYBQ/MemoSync.git
cd MemoSync
bun install
bun run build
bun run start

Desktop development and packaging are macOS-only and target the host architecture:

bun run desktop:install
bun run desktop:dev
bun run desktop:package:mac

How do you use this agent?

Open Settings → Models, pick an engine (Claude Code, Codex, Pi or OpenCode) and configure access: reuse your existing sign-in, or click Add connection to choose a built-in provider or a Custom endpoint, press Test to send one small request, then pick a default model. If the engine shows Not signed in, sign in from your terminal and press Detect again.

Then click + in the sidebar → Add project, open a folder and send a task; a fresh library starts empty, so early turns are mostly proposals to review. To serve it elsewhere:

bun run start --port 4000
bun run start --no-open
bun run start --remote
bun run start --password <pw>

Memory behavior is set under Settings → Memory: Use Memory, Learn From Chats, Control (Review each turn vs Automatic), Memory Changes and Memory Preview. Data export and development commands:

bun run export-data
bun run export-data --full
bun run dev
bun test src/

What are this agent's strengths and limitations?

Pros
  • Memory is broken into scoped, versioned, historied items that must pass human review before being written, which addresses the unbounded, unauditable growth of CLAUDE.md.
  • The audit runs separately from the agent's own claims, so a violation the agent never cited still surfaces, with Where used linking back to the exact sentence and tool call.
  • Four engines are supported (Claude Code, Codex, Pi, OpenCode), and OpenAI-Chat-only connections are translated through a local 127.0.0.1 gateway so the engine never sees your key.
  • All state stays in ~/.memosync/data/ with API keys in a user-only-readable credentials.json, and --password, --remote and --share cover self-hosted access.
  • An Automatic mode plus one-click Revert balances strict per-turn review against quick questions.
Limitations
  • The official DMG targets Apple silicon Macs only and is not notarized, so the first launch needs a manual security override.
  • It depends on externally installed agents: Codex must be present, Pi needs a specific package version, OpenCode needs 1.17.4+, and stale versions rely on the engine card's Upgrade button.
  • Audit verdicts are model judgments that the README itself says can be wrong, so they are not compliance evidence.
  • The app and a source install share ~/.memosync/data and cannot run at the same time.
  • Conversations, selected memories and project context are still sent to your model provider; this is not fully offline memory.

How does this agent compare with similar options?

Against coding agents that rely on static memory files such as CLAUDE.md, MemoSync turns memory into versioned, reviewable, auditable items and adds violation detection plus mid-turn interruption; the cost is a separate local server and a per-turn confirmation loop. The README notes the project started as a fork of Kanna, useful context for its lineage.

Key facts side by side with the most closely related agents.

Agent Source review Form / cost Stars Updated Language Full support on
MemoSync This agent 50 · Major gaps Desktop appFree + model costs ★ 396 6d ago TypeScript Codex · Claude Code · OpenAI API · Claude API
Emulo 76 · Good Agent plugin / skillFree + model costs ★ 306 1mo ago HTML Codex · Claude Code
Solo Agent 55 · Major gaps Self-hosted serviceFree + model costs ★ 605 2mo ago Go Codex · Claude Code
PRO-LONG 60 · Some gaps CLIFree ★ 465 18d ago Python Codex · Claude Code

How does FollowAgents rate this agent?

FollowAgents source review · FARS-2.1
Major gaps
50/ 100 5-point scale 2.5 / 5
This review covers the source as of Nov 11, 2026. The project has shipped 2 stable release(s) since, which have not been re-reviewed. The score and verdict describe the reviewed version only. Recent releases ↓
Trust 16/29
Reliability 6/14
Adaptability 9/18
Convention 8/18
Effectiveness 7/13
Verifiability 4/8
Why each dimension lost points
Trust16 / 29 · 2.8/5

README honestly states the agent and its terminals run as the user with filesystem access and that profile isolation separates configuration, not the filesystem — transparent but a large privilege surface, so least_privilege is 1. User confirmation is consistently documented and reflected in code (per-turn review, Ask First, sensitive items always ask, Stop on citation), scoring 2. Data flow is clearly described (local ~/.memosync/data, nothing uploaded, provider receives conversation and memories, imports are read-only), scoring 2. Sensitive handling has looksLikeSecret and exclusion of sensitive items from the plaintext Markdown projection, scoring 2. The dependency set is large and includes beta pins (@legendapp/list beta, cloudflared, several SDKs) with no audit or pinning policy shown, so dependency_security is 1. External effects exist via --remote/--share/cloudflared with only documentation-level warnings, scoring 1. Rollback is supported by Revert, archive, Removed restore and SQLite as canonical source, scoring 2. Source attribution credits the Kanna fork and third-party notices in LICENSE and Acknowledgements, scoring 2.

Reliability6 / 14 · 2.1/5

Code comments and README agree that SQLite is canonical and Markdown is a generated export rebuilt on startup, scoring 2. Dependency availability hinges on external CLIs (claude/codex/opencode/pi) and version floors; documented but with no fallback or verification mechanism, scoring 1. Failure messages are thin — mostly console.warn and a few throws, with little user-facing error handling evidence, scoring 1.

Adaptability9 / 18 · 2.5/5

Covers multiple engines (Claude Code, Codex, Pi, OpenCode) and two control modes, scoring 2. Capability boundaries are stated (no bundled agents, macOS-only desktop, Windows unsupported, audit can be wrong), scoring 2. Trigger precision is weak: memory injection defaults to all-in and citation/audit depend on model judgment, scoring 1. Environment fit is narrow — macOS packaging only, Linux from source, Intel Macs need source, scoring 1.

Convention8 / 18 · 2.2/5

README is well structured and sectioned, scoring 2. Install notes cover DMG, source, Bun version and the notarization workaround, scoring 2. Naming stability is marred by Kanna leftovers (KANNA_RUN_LIVE_TITLE_TESTS, Kanna test comments in CI), scoring 1. Examples are mostly screenshots and video with no textual FAQ or troubleshooting, scoring 1. Known limitations are scattered (not notarized, audit may be wrong, single instance) rather than consolidated, scoring 1. LICENSE is full MIT preserving upstream copyright, scoring 2. Versioning has only a package.json version with no CHANGELOG, scoring 1. Maintenance responsibility is unclear: publisher unverified, no maintainer or support channel stated, scoring 1.

Effectiveness7 / 13 · 2.7/5

Output usability is concrete — review cards, citations, audit verdicts and a Memory Board with a defined data model, scoring 2. Marginal value is real: it makes implicit agent memory explicit and auditable beyond native behavior, scoring 2. Cost/benefit is weaker: per-turn review is heavy, audit and refinement consume model calls, and users must supply keys or sign-ins, scoring 1.

Verifiability4 / 8 · 2.5/5

Claim traceability exists via citations and audit links to sentences and tool calls, but the README itself says citations are not proof, scoring 1. Cross-source corroboration is limited to README, AGENTS.md, a few source files and CI configs, scoring 1. Fact/inference separation is good: the README distinguishes agent-reported from audit-found verdicts and admits the audit can be wrong, scoring 2.

Risks and how to mitigate them
  • The agent and its terminals run as your user with filesystem access; profile isolation separates configuration, not the filesystem — assess the privilege surface before use on untrusted projects.
  • --remote, --share and cloudflared can expose the local server to LAN or the public internet; avoid by default and pair with --password when enabled.
  • The dependency set is large, includes beta pins and external CLI version floors, with no stated audit or pinning policy.
  • Memory injection defaults to all-in and citation/audit depend on model judgment; the README admits the audit can be wrong, so it is not a compliance proof.
  • Publisher identity is unverified and Kanna naming/test configuration remains in the code, leaving maintenance responsibility and update path unclear.
  • The macOS app is not notarized and requires bypassing Gatekeeper on first launch; verify the download source.
Evidence confidence: Low Reviewed Nov 11, 2026 Reviewed revision e16321576a2e New commits since this review; the score may not cover them
See the full review method →
Maintain this project? Add the FARS badge to your README

The badge shows the current review score and verdict and links back to this page. It refreshes about once an hour after a re-review.

FARS

[![FARS review](https://followagents.com/badge/memosync.svg)](https://followagents.com/en/agents/memosync)
<a href="https://followagents.com/en/agents/memosync"><img src="https://followagents.com/badge/memosync.svg" alt="FARS review"></a>

Recent releases

Latest entries from the project's GitHub Releases. Notes are quoted from upstream as written, not rewritten or tested by FollowAgents.

  1. v0.45.0 MemoSync v0.45.0
    Release notes excerpt
    MemoSync v0.45.0 brings the desktop app to Windows, Linux and Intel Macs, alongside Apple silicon. Desktop platforms - Windows x64: a per-user installer and a ZIP containing MemoSync.exe. - Linux x64: an AppImage and a Debian/Ubuntu .deb package. - macOS: separate DMG and ZIP downloads for Apple silicon (arm64) and Intel (x64). …
    View on GitHub ↗
  2. v0.44.0 MemoSync v0.44.0
    Release notes excerpt
    MemoSync v0.44.0 adds controls for the full memory lifecycle: retire rules, correct audits, set expiry dates, detect file changes, and share project rules through Git. What's new - Archive from where you work. Retire or restore a memory from Working Memory, citation cards, audit rows, or the Memory Board. “Open in Board” opens the cited memory without leaving your chat or losing its draft. …
    View on GitHub ↗
  3. v0.43.1 MemoSync v0.43.1
    Release notes excerpt
    MemoSync v0.43.1 fixes Codex chats after a permission change and two composer glitches. Fixes - Codex after a permission change. In a new Codex chat, switching the mode (for example to Full access) before the first message made every send fail with no rollout found for thread id …. MemoSync now starts a fresh Codex thread in that case; the old one had never run a turn, so nothing is lost. Chats stuck this way recover on their next message. …
    View on GitHub ↗

FAQ

Does MemoSync call models itself, and what does it cost?
It does. The software is free under MIT, but splitting memory into items, auditing, and Automatic-mode clean-up all need a model, billed through your own sign-in or API key.
Can it read my existing CLAUDE.md and Codex memory, and does it modify them?
Yes. Import memory reads the Claude Code and Codex memory folders and has your chosen model split them into personal-preference and project items; nothing is saved before review, and the import does not change the original folders.
What if an audit verdict is wrong?
The README states the audit is a model judgment that can be wrong. Every verdict is labeled as found by the audit or reported by the agent, and Where used jumps to the source sentence and tool call so you can check it.
Where do my data and keys live?
Everything is under ~/.memosync/data/: SQLite memories with the Markdown copy, chats and settings, API connections, and a local usage log you can disable with MEMOSYNC_USAGE_LOG=0. Keys sit in credentials.json, readable only by you and never sent back to the browser.
Can I reach it from another machine or over the LAN?
Yes: bun run start --remote listens on 0.0.0.0, --password requires a password before the app loads, and --share creates a temporary public URL with a QR code. Note the agent and its terminals run as your user with filesystem access; profile isolation separates configuration, not files.
Is MemoSync open source?
Yes. The source code is on GitHub (SLEEPYBQ/MemoSync) under the MIT licence.
Is MemoSync safe to use?
FollowAgents' FARS review (Nov 11, 2026) scored MemoSync 16/29 on trust and safety. None of the eight safety controls the review checks for was missing outright. The repository has changed since that review.
Is MemoSync still maintained?
Yes. The latest commit on GitHub is from 2026-11-15, within the last 30 days. The latest release is v0.45.0 (2026-11-14).
Who makes MemoSync?
MemoSync is published on GitHub by the account SLEEPYBQ. FollowAgents has not verified which person or company is behind that account.
What is MemoSync's official website?
The project's GitHub repository lists https://memosync.cc/ as its website. That link is set by the repository owner; the source code itself is on GitHub (SLEEPYBQ/MemoSync).
Was this page helpful?
View on GitHub ↗ Install ↓

Host ecosystems

Compare agents like this one

The same FARS review applied across the shortlist this agent qualifies for.

Related concepts