Your agents need memory
Artur Sabirov1 min readBrain launched today: one shared memory for your agents, readable and writable over MCP, backed up on schedule, scoped per product. It fixes a waste we kept paying without noticing.
An agent starts every session blank. It re-learns your company from whatever someone pastes into the prompt, performs at the level of that paste, and forgets all of it at the end. Meanwhile the real context, the decisions and constraints and history, sits in tasks and docs and code where the agent can't see it. Multiply that by every session, every agent, every day. We paid that tax for months and barely felt it, because each paste felt small on its own.
Brain lets agents store facts and the relations between them, then search both: "what do we know about billing," and "what's connected to that decision." Memories carry scopes, a shared space the whole team's agents read and private spaces per person, so an agent can tell company knowledge from one person's notes. The operational parts are the platform's usual habits: backups run on schedule, restore is an operation you run rather than a ticket you file, and memory is scoped per product so work on product A doesn't leak into product B. Stats show what's accumulating, so the memory doesn't quietly rot into a junk drawer nobody can audit.
Here's the rule this launch won't break. A memory your agents share shapes every answer they give, so a memory filled invisibly is a risk you can't inspect. Access follows the platform rule, granted and named and revocable, the keys to one room and not the building. An agent writes memory under its own grant, so every remembered fact is attributable to the session that stored it. Every entry keeps its source.
Knowing what your agent knows is the same promise as knowing what it can touch. List what your agents remember, trace where each fact came from, delete what shouldn't be there. All three, today.
Related posts
Building Lessly in the open
Follow along, and get early access to the private beta.