Project Brain overview
How Capreolus stores one long-lived agent per project with vector memory.
One agent per project
Most agent products spin up a fresh assistant for every chat. Capreolus does the opposite: a single long-lived agent — the Project Brain — lives for the lifetime of a project and accumulates context. That means the agent that drafts your changelog already knows what shipped, who reviewed it, and which tickets it closed.
Vector memory
The Project Brain stores everything it has read or written as embeddings in a project-scoped vector store. Searches are similarity + recency weighted, so a question about "the iOS crash from last Friday" pulls the relevant Sentry issues even if you describe them with different words.
What gets indexed
- Release notes and version metadata.
- Sentry / Firebase issues and their resolutions.
- Customer feedback (from the in-app prompt and store reviews).
- Documents you upload — design docs, RFCs, runbooks.
- Past agent runs (for retrieval-augmented prompting).
Privacy boundary
Memory is hard-scoped to a project. The agent in Project A cannot read anything from Project B, even if you own both. We do not train third-party models on your project data; embeddings are generated locally and used only for retrieval.
Where to go next
- Browse the available skills: Skills catalog.
- Understand how memory is stored and ranked: Memory model.
- Programmatic access lives under
POST /v1/agents/{projectId}/runs.