brainbox
brainbox
Memory as a service — personal preview

One memory, carried across the AI tools you connect.

brainbox keeps the context that matters — decisions, constraints, lessons learned — and hands it to the assistants and agents you've connected it to. Search what's saved, correct it when it changes, forget it when it stops being true.

Currently a personal preview — invitations open gradually, and registering interest doesn't promise a date or immediate access. The example below is synthetic and runs entirely in your browser: no model calls, no network, nothing you type leaves this page.

Watch context travel. Then take it back.

A walkthrough of the four things brainbox does — remember, recall, correct, forget. Write a memory in one client, recall it from another, fix it when it changes, remove it when it's wrong. Provenance rides along with every step.

Synthetic example Everything here is fabricated for the walkthrough — no model, no network, nothing you type leaves this page.
  1. 1 remember
  2. 2 recall
  3. 3 correct
  4. 4 forget

Claude Code

client A — writes

End of a working session. Ask the agent to keep the one constraint that keeps coming back:

ChatGPT

client B — recalls

Next morning, a different tool — same you. It already knows what matters, and says so when it doesn't:

ChatGPT

client B — corrects

The rule changed this afternoon. Fix the memory in place, from the client you're already in:

your workspace

owner — forgets

It stopped being true at all. Forgetting is an owner action — deliberate and recorded, never done silently by a machine.

About the labels: recall can come back weak (saved as unconfirmed — verify before relying on it) or unknown (nothing confidently saved — brainbox says so instead of guessing).

Connect once. Then it's just there.

brainbox is a private memory service your clients connect to. Once a client is configured to use it, your AI can recall what's relevant and remember what matters — always explicitly, never silently behind your back.

01
Connect

Pick your client and follow its setup guide — Claude, ChatGPT, Codex, opencode, goose, Amp and more. Sign in with Google once; apps you approve get a key of their own.

02
Work normally

Once your client is set up to use it, your AI can recall what's relevant before it answers, and remember what matters when asked. Nothing is remembered silently in the dark — lessons are never invented behind your back.

03
Stay in control

Search by meaning or keyword, revisit recent saves, correct in place, and forget for good. Deletions are recorded, and recovery checks the current deletion record before restoring anything.

Twelve clients. Twelve guides. No mystery.

brainbox connects over an open protocol, so assistants and agents that support that kind of connection can use it. Setup guides exist for twelve clients today — and every guide states exactly how far it's confirmed. Guides are guides, not a claim that every integration is freshly verified.

Recipe-verified setup guides
Claude — web & desktop Claude Code ChatGPT Codex CLI opencode goose Amp
Legacy & unverified recipes — re-verify before relying on them
Antigravity IDE legacy OpenClaw legacy Hermes Agent legacy Grok unverified
Not available yet
Gemini
Anything else that supports the same open connection standard
Your own agents & scripts

ChatGPT is the client connected to the founder's own workspace today; the other verified recipes are confirmed against each app's own documentation, and each guide says exactly what was tested.

Available today, planned next, and current limits.

What works right now, what's planned next, and the current limits.

Available in the current personal preview
  • Sign in and connect clients — sign in once, approve each client yourself, revoke it any time.
  • Remember and recall — save what matters (events and explicit lessons), then find it again by meaning or keyword, or browse recent saves.
  • Correct and forget — fix a memory in place when it changes; remove it for good when it's wrong.
  • Graph & Journeys — threads that connect related memories. Similarity can suggest related items; a relationship becomes real only when you confirm it — relatedness never quietly becomes lineage.
  • Your memory is isolated — a separate database of your own, not shared tables with anyone else.
  • Snapshots you can recover from — daily backups kept two weeks; recovery is deliberate: it checks the current deletion record and refuses to guess.
  • One private workspace — a personal preview today; no public signup, no shared deployment.
Current limits
  • This is a fresh rebuild. Memories from the original brainbox haven't been imported yet — the original storage is awaiting recovery, and not every old feature has been migrated.
  • No promise of perfect durability. What we do instead: daily snapshots, two-week retention, and a recovery process that checks the current deletion record and refuses to guess.
Why not just ChatGPT memory?
Their memory lives in their walls.
Yours should live with you.

Built-in chatbot memory is real — and it stops at one vendor's door. brainbox is one memory across the clients you connect: the one you use today, the one you'll switch to, and the scripts you write yourself. Correct it, forget it — and if you ever need it back, an operator-assisted recovery process restores your snapshots rather than leaving your history locked inside one chat app.

Register interest

Personal preview today — invitations open gradually.

Built by someone who needed it first.

brainbox is a project by Barak Achillah Asidi, founder of CIF AI — a small studio in Kenya. He's a data engineer by trade; other CIF AI experiments include Tarjumi and Ishara.

brainbox didn't start as a product. It started as the memory behind his own AI, and became a product because working without it stopped making sense. What you read about above is the fresh rebuild he uses himself.