MemBridge reads the session logs your Claude Code and Codex agents already write, distills what each one worked out, and gets it to everyone else's agents. At startup, mid-session, and on demand.
One command, about two minutes, no account. npm i -g @membridgeai/membridge
Most memory tools stop at writing a file. The useful part is what happens after the session has already started, and months later when nobody remembers who did it.
We measured how much this costs: across 6,104 real agent sessions from 310 developers, the median developer spends 31% of their tokens re-orienting an agent in code it has already seen. Read the Context Ledger →
What the team's agents worked out gets written into the instruction files your agents already read, attributed by person. Nobody writes a doc. Nobody re-explains the architecture.
Your agent opens a file a teammate's agent made a call about, and that call enters the running session. Once, exactly when it matters. No doc, no Slack, no human relaying it.
A ranked search over what every agent on the team has actually done, exposed over MCP so any agent can query it directly. The useful answer is usually months old.
Not keyword matching over a digest. It ranks across the decisions people made, the gotchas they hit, what they were trying to do and the files they touched, filterable by teammate, project, file or date, over your team's full synced history.
Ranked by relevance, not recency. Work from months ago surfaces when it answers the question.
Every agent session across your team, who can see which project, and where knowledge is concentrated in one person's head.
Everything below is a mock of the shipped app, filled with sample data. The numbers are illustrative, not measured. Every screen in the sidebar is in the app today. Click through it.
| Project | Sessions · 7d | Sync | Priya | Marcus | Jules | Sam |
|---|---|---|---|---|---|---|
| api-gateway | 14 | ✓ up to date | ✓ | ✓ | ✓ | ✓ |
| web | 9 | ✓ up to date | ✓ | ✓ | ✓ | · |
| billing | 6 | ✓ up to date | · | · | ✓ | ✓ |
| infra-secrets Private | 4 | not shared | – | – | – | – |
Tick a cell to grant or revoke, without opening the project. A private project cannot be granted until it is shared. Every change lands in the audit trail.
One memory across both, so switching tools mid-task doesn't reset anything.
Your sessions, summaries and code leave this machine only as what you have explicitly shared with your team, encrypted.
In the real app, live rows show the ask a session opened with, never a guess at what it is doing.
Cross-tool memory. Claude Code and Codex share one store. Switching tools doesn't reset context.
Auto-registers over MCP. Finds the agents you have installed and wires itself in.
Ask why about any file. Which session touched this line, by whom, and the reasoning behind it.
Per-project access control. Pick exactly who sees which project, changed in one grid.
One-command onboarding. A new teammate joins with a single invite link.
Roles and an audit trail. Owner, admin, member. Every membership and sharing change recorded.
Health at a glance. Who's syncing, whose keys are verified, and anything that quietly stopped reaching the team.
Runs local. A daemon on each machine. No agent of ours in your CI, no repo upload.
Team sync is encrypted on by default. Keys are generated on each laptop and never leave it, so the backend stores ciphertext it cannot open. The content is what gets encrypted; routing details (author, project, timestamp, tool) stay readable so sync and access control can work.
Summaries and file paths sync. Your verbatim prompts stay on your machine unless you opt in, per session or per project.
Removing someone mints a new team key sealed only to the people who remain. Secrets are stripped before anything is written or sent.
Install on one repo and watch a session get distilled. Team sync is a separate, deliberate step.
Desktop app for macOS (Apple Silicon) and Windows. On Linux or an Intel Mac, the npm command runs the same daemon and serves the same dashboard in your browser.
Node 18 or newer. The desktop app adds a menu bar or system tray icon and start at login. Everything else it does, the CLI does too.
Free while we're early. No account, no card, through v1.0. When we charge it will be per seat for team sync. Individual use stays free.
MemBridge is early. If you want to see what breaks before you standardise on it, you are exactly who we want to hear from.
<!-- membridge --> markers in CLAUDE.md and AGENTS.md. Your own notes are
never touched, and membridge remove restores every file exactly.No. MemBridge reads your agent sessions locally and writes memory to disk on the same machine. Your sessions, summaries and code go nowhere until you turn on team sync, and even then it is only what your sessions concluded: decisions, summaries, and file paths, never your source code. Team sync is end-to-end encrypted, so the backend stores ciphertext it has no key for. Separate from your work content, MemBridge sends anonymous usage counters and diagnostics tied to a random install id (turn them off with the diagnostics setting or MEMBRIDGE_NO_DIAGNOSTICS=1), and it asks GitHub about once every six hours whether a newer release exists.
They are stripped before anything is written to disk or sent to your team. MemBridge matches the shapes of AWS, GitHub, Google, Slack, OpenAI and Anthropic keys, JWTs, PEM blocks, connection strings with inline credentials, and bearer tokens, plus a high-entropy catch-all for the ones nobody has a pattern for. Redaction runs at every boundary, including on data arriving back from your teammates.
A hand-written context file is the startup half of this, and it only stays useful while somebody keeps editing it. MemBridge writes that file for you from what your sessions actually concluded. The two things it does that a file cannot: deliver a teammate's decision into a session that is already running, and answer a question against months of your team's history that no file would still contain.
Claude Code and Codex are first class. Their sessions are what MemBridge reads, and they share one memory, so switching between them mid-task doesn't reset anything.
Cursor is different, and the distinction matters if your team is split. MemBridge registers itself over MCP with Cursor, so Cursor can query everything the team knows. It does not read Cursor's sessions yet, so work done in Cursor does not feed back into the shared memory. Cursor consumes, it does not yet contribute.
Removing them revokes access to every shared project immediately and mints a new team encryption key sealed only to the people who remain, so their copy stops being able to read anything new. Membership and sharing changes are all recorded in an audit trail.
Yes, at three levels. Verbatim prompts stay on your machine by default; only summaries and file paths sync unless someone opts in. Any project can be kept private, and shared projects have per-person access. A developer can pause capture entirely for a project without affecting anyone else.
No. It watches the sessions your team already runs and writes to the instruction files your agents already read. Nobody writes documentation for it, and nobody has to remember to record anything.
Send us a note and we'll set it up on a repository your team actually works in, then show the hand-off firing between two real sessions. No slides, and nothing to uninstall afterwards if it isn't for you.
Useful to know beforehand: how many engineers, which agents they run, and whether your repos live on GitHub.