Models kept getting better.
Their memory did not.
Every frontier release buys more reasoning per session. None of it survives the session. Re-orienting an agent that remembers nothing typically burns 30,000 to 150,000 tokens before it does useful work, and the first thing a context window loses is the negative knowledge: the paths already ruled out. supabrain is the layer that keeps it.
Minutes to connect. Every session after that starts informed.
Install the app, point your agent at it over MCP, and the store builds itself out of the work you were doing anyway. No stack to stand up, no daemon to babysit.
It stores conclusions, not transcripts.
Two things reach the store: what an agent verified during real work, and the curated docs that map a codebase. Both are checked at write time, and both stay searchable.
The answer is worth keeping. The transcript is not.
A chat log is a haystack. supabrain keeps the needle: the decision that was made, the fix that worked, and the option that was ruled out. Each one is validated at write time, given a stable ID, and deduplicated against what is already there, so the store stays clean enough to trust.
One query goes in.
The right conclusion comes back.
Full-text search always runs. On-device semantic search adds meaning-based recall, and an optional re-ranker sharpens the top results. Nothing is sent anywhere to make that work, so a cold laptop on a plane still answers.
It remembers the answers that were traps
The most expensive thing an agent does is confidently re-walk a path you already proved wrong. Ruled-out options are first-class memories here, so the next agent is warned instead of repeating the week you already spent.
Docs start lying the week after you write them.
Code moves and documentation does not. supabrain generates a codebase's docs once, then flags the exact pages that drifted as the code changes, and fixes them for pennies with a receipt on every run.
The episodic why, and the current map
Memories record what happened and why it was decided. Curated docs record how the system works right now. One search call fans out across both.
Saved once at 09:41.
Already paying off at 09:42.
A memory is written once and recalled for as long as it stays true. That is the whole economic argument: the store is the only part of your AI spend that appreciates.
PLATFORM-APIOne connection. Any agent tool.
supabrain speaks the Model Context Protocol, so saving and recalling are ordinary tool calls. The server is a stdio process your client spawns on demand: there is nothing to start, no daemon to babysit, and no cloud round-trip to make search work.
Everything it takes to make memory trustworthy
A memory store is only worth having if you can trust what comes back out of it.
One store, widening from the developer outward
Memory you own, not memory you rent
The mechanism is not the moat. Setup, ownership, and portability are.
Flat pricing. No usage meter.
Search and embeddings run on your machine, so serving you costs us almost nothing and the price does not climb with your store. Free local tier, cloud sync at 10 USD per user per month, teams at a flat 25 USD per person, enterprise self-hosted on request. Billing opens with general availability; early-access accounts are granted without payment.
Common questions
What is supabrain?
supabrain is persistent, local-first memory for AI agents: one brain across every tool, every session, every machine. It captures the decisions, fixes, and dead ends your agents produce, then lets any later agent recall the right one in a single query.
How is this different from a long context window?
A context window is rented for one session and lost at the end of it. supabrain stores conclusions outside the model, so they survive the session, the tool, and the model. It also keeps negative knowledge, the paths already ruled out, which is the first thing a context window drops.
Which tools does it work with?
Any client that speaks the Model Context Protocol. Claude Code, Cursor, and Gemini CLI are verified, and any other MCP-compatible client connects the same way. The server is a stdio process your client spawns on demand, so there is nothing to run.
Is my data private?
Your memories live in a single database on your machine, and storage, keyword search, and semantic search all run on-device. Embeddings are computed locally and are never sent to the cloud, even with sync on. Data leaves only when you turn on a feature that needs the network, such as sync or an LLM-backed cleanup pass.
What stops the store filling up with noise?
Every save is checked at write time against a format gate that returns a clear verdict, so a malformed memory is bounced rather than silently stored. Near-duplicates are merged and outdated entries superseded, and because nothing is hard-deleted, every such change is reversible.
What happens when I switch models?
Nothing. Memory lives outside the model, so changing model or tool costs you nothing and the store comes with you. That portability is the point: the store is the part of your AI spend you keep.
What is the docs feature?
supabrain generates a codebase's curated docs once, then flags the exact pages whose source has drifted as the code moves, and fixes them on a cost ladder with a committed receipt on every run. Detection is free. In a measured run in July 2026, a pass moved pointer accuracy from 82% to 99% and cut drifted pages from 30 to 1, at roughly 0.45 USD per page for the routine tier. Humans still decide what deserves a page.
Can I buy it today?
Not yet. supabrain is in early access with the beta open, and all tiers are functional, but checkout and billing are not live. Early-access accounts are granted without payment. The tiers are Free for local use, 10 USD per user per month for cloud sync and multiple devices, a flat 25 USD per person per month for teams, and a self-hosted enterprise option on request.
Who builds supabrain?
supabrain is built by tec.tech, a software studio in La Réunion that builds its own products rather than reselling other people's.