You explain your architecture to Claude Code on Monday. Tuesday you explain it again to Codex. Every agent starts from zero, every session, and the decisions you already made evaporate when the window closes.
MemBridge watches your sessions and writes what matters into the files every tool already reads. The next agent you open already knows. 100% on your machine.
Intel Mac or Linux: npm install -g @membridgeai/membridge
And the part git blame can't do: ask any line why it exists.
Sam makes a call in Claude Code tonight. MemBridge writes it into the shared memory both AI tools read, CLAUDE.md and AGENTS.md. When Maya opens Codex tomorrow, it already knows.
Your team did things while you were gone. The dashboard shows you exactly what: project by project, session by session, down to the word-for-word prompts.
All caught up across your projects. Click one to open it.
Checkout copy swap: "Reserve your spot", mobile spacing pass
Empty-cart rework, config cleanup, retry hardening
Coupon stacking rules pinned down before the sale
2 members · created July 2026 · you are a member
Anyone with this code joins as a member. Regenerating invalidates the old code.
You're a member. Only the owner can rename the team, change roles, or manage the join code.
Watched projects, team, account.
Watched projectsShips with the daemon: membridge dashboard opens it. Menu-bar app on macOS.
MemBridge is built on a different bet: the memory should capture itself, stay local, and live in the files your tools already read. Here's what that buys you.
The daemon tails your Claude Code and Codex sessions and distills them on its own. No "remember this" tool to call, no notes to maintain. Forgetting to save is the thing that kills every other memory tool.
It runs where your code already lives. Binds to 127.0.0.1, zero network calls in solo use, no account, no telemetry. Nothing to trust because nothing leaves.
Memory is scoped to each repo, not one global blob. What you decided in cartful never leaks into the next project's context.
membridge why file:line surfaces the prompt and decision behind any line of code. git blame tells you who; MemBridge tells you why. No other memory tool does this.
Claude Code and Codex are built in; anything that logs JSONL plugs in via config. The neutral bridge no single vendor will build across its rivals.
Memory lands between markers in CLAUDE.md / AGENTS.md: readable, diffable, and membridge remove restores everything byte-for-byte. No lock-in, no black box.
Invite your team to a project and every Claude Code and Codex session lands in one shared feed: the plain-English summary up top, the exact prompts one click down. End-to-end encrypted, invite-only, and off by default.
Projects stay private until you share them. Only the people you invite can see a project's memory, and anyone can pause or leave at any time.
Your team's shared memory is encrypted client-side before it syncs. The relay only ever moves ciphertext: it can never read your code, your prompts, or your summaries.
Catch up on any teammate's AI coding session in seconds: what changed in plain English, plus the word-for-word prompts that produced it.
No accounts, no cloud, no setup ritual. MemBridge reads the session logs Claude Code and Codex already keep, and writes to the context files they already read.
membridge start watches your Claude Code and Codex transcripts: read-only, incremental, byte offsets only. Any tool that logs JSONL plugs in with a config entry. No code.
Recent prompts, files touched, and the decisions behind them become a short per-project digest. Secrets (sk-…, AKIA…, ghp_…) are scrubbed before a single byte is written.
The digest lands between <!-- membridge --> markers in CLAUDE.md and AGENTS.md. Your own notes stay byte-for-byte intact, and membridge remove restores everything exactly.
Your code never leaves your machine. Not as telemetry, not as analytics, not as an oops. MemBridge is local-first by architecture, and every claim below is one you can check yourself.
The daemon binds to localhost and nothing else. Solo use makes zero network calls: no telemetry, no analytics, no accounts. Turn your wifi off and MemBridge won't notice. Team sync is opt-in, off by default, and end-to-end encrypted.
The core is plain Node with zero npm dependencies. No supply chain to attack, no transitive packages to audit, nothing to read but the code itself. And the code is on GitHub.
API keys and tokens (sk-…, AKIA…, ghp_…) are scrubbed before a single byte hits a context file. Redaction happens before write, never after. Add your own patterns in config.
MemBridge only ever writes between its own markers, and never touches your code. One command, membridge remove, strips every trace and restores your files byte-for-byte. Leaving takes ten seconds.
With MemBridge running: one socket. It's your dashboard, and it only answers localhost.
Menu-bar app plus the membridge CLI. Signed & notarized — double-click and it opens, no warnings, no terminal.
Download for MacOther platforms: Windows · Intel Mac & Linux run npm install -g @membridgeai/membridge
Clone it, read it, run it. Zero dependencies, so what you read is what runs. Your dashboard comes up at 127.0.0.1:7437.
Occasional updates. Nothing else, ever.
Run membridge start. The daemon watches your Claude Code sessions, distills each one into a per-project memory, and writes it into CLAUDE.md, which Claude Code reads on every startup. Your next session starts already knowing what the last one did.
Persist the context somewhere the tool re-reads on startup. That's exactly what MemBridge automates: it distills every session into a per-project memory and syncs it into CLAUDE.md and AGENTS.md, so a new session, even in a different tool, picks up where the last one left off instead of starting from zero.
The same memory block lands in both CLAUDE.md and AGENTS.md, the files each tool reads natively. Codex sees what Claude Code did and vice versa, no copy-pasting, no MCP wiring.
No. The daemon binds to 127.0.0.1 and solo use makes zero network calls: no accounts, no telemetry, nothing phoning home. You can verify it yourself with one lsof command, shown in the Privacy section above. Team sync is opt-in and end-to-end encrypted.
Never. MemBridge only writes between its own <!-- membridge --> markers, your notes stay byte-for-byte intact, and membridge remove restores every file exactly. It's plain markdown you can read and diff.
CLAUDE.md is the markdown file Claude Code automatically reads at the start of every session, a project's standing instructions and memory. MemBridge keeps a distilled, always-current summary of your recent AI sessions inside a marked block in that file, so Claude Code starts each session already caught up.
AGENTS.md is an open convention for giving coding agents per-repository instructions. Codex and a growing list of AI tools read it natively. MemBridge writes the same shared memory block there that it writes to CLAUDE.md, which is how context crosses between tools.
Claude Code and Codex adapters ship built in. Any tool that logs its sessions as JSONL can be wired up with a small config entry, no code required. More first-class adapters are on the roadmap.
Four ways. MemBridge captures automatically by tailing session logs, so there's no save tool for the agent to forget. Memory is per-project, not one global store that bleeds between repos. It runs 100% locally with no API key or cloud service. And membridge why file:line answers a question git blame can't: the reasoning behind any line of code.
Yes. MemBridge is free and source-available with zero dependencies. There is no API key to bring and no paid tier. Install it and you have all of it.
Invite teammates to a project and every Claude Code and Codex session lands in one shared feed: plain-English summaries plus the exact prompts. End-to-end encrypted before it leaves each machine, invite-only, off by default.
Each can build memory inside its own tool. Neither has a reason to build the bridge across its competitor's, and that neutral spot is exactly where MemBridge lives. The cross-tool provenance it accumulates is the part that compounds.
The Windows build isn't code-signed yet — an Authenticode certificate is in progress — so SmartScreen shows a caution prompt the first time you run it. Click More info, then Run anyway. The macOS build is fully signed and notarized, so it opens with no warning. Prefer no prompt at all? npm install -g @membridgeai/membridge installs the same thing with none.
Still reading? The repo answers the rest.