Braintrust
A Slack-like workspace where agents share memory and find what they need.
The question I was chasing
What does it feel like to work with a team of agents, not just prompt one?
A single chat forgets everything between sessions. I wanted agents that share context the way a team shares a Slack workspace: channels, files, and a memory they all draw on.
The constraints
Agents need the right project context without stuffing every past message into the next reply (the short memory window every model carries). That constraint is the whole design problem: memory and search have to sit at the centre, not the edge. Get that wrong and the agents either forget what matters or drown in what does not.
The decisions that mattered
I built the shared memory in from the start, not bolted on later. Agents find related notes by meaning across projects, and that memory sticks between sessions. That is what the workspace is made of.
The agent builder supports both OpenAI and Anthropic, instead of locking the team to one provider.
I kept a Slack-like model, with channels and shared context, because it is an instantly familiar picture of several people collaborating, which is exactly what a team of agents is.
Where it landed
A working prototype, and the shared memory does what I hoped: agents reach for the same project context without being handed it. Coordination is still open: nothing in it decides which agent takes which job.
Part of the Rolling Waves work archive.