Pragmus is a managed exo-brain for teams working with AI agents. It captures decisions from agent output and human-AI sessions, structures them, and makes them readable by any agent via MCP.
the problem
Architecture calls, API contracts, naming conventions — made in one session, forgotten by the next. Your agent re-proposes what you already rejected.
Three agents working on the same project. None of them know what the others decided. You become the bottleneck, re-explaining everything.
Manual files work for one person, one repo. They break the moment a project spans multiple agents, sessions, or contributors.
Last month’s directive contradicts this week’s feedback. Nothing connects them. Nobody knows which one is current. The project drifts.
structured knowledge
Top-level boundaries for your project — architecture, product, operations, whatever maps to how your team thinks. You set these. Everything else organizes inside them.
Within each area, topics form automatically as related knowledge accumulates. No folders to create — the structure reflects what your project is actually about, and it evolves as the project does.
When items relate to each other — a decision followed by progress, a constraint that contradicts an earlier call — they form threads. Contradictions surface. Progress is tracked. Nothing is orphaned.
The smallest piece — a single decision, constraint, observation, or direction. Discrete, classified, queryable. Not a blob of notes. Something an agent can read and act on.
Summaries are generated and maintained at every level — always reflecting the latest state. Agents read the level they need.
what goes in
Whether it’s an agent completing a task or a human thinking through a problem with AI — the decisions that emerge are captured as they land. The next session starts from what was already resolved, not from scratch.
how agents read it
This isn’t something you sit down and read. It’s a knowledge graph — multi-dimensional, interconnected, always current. Agents traverse it at runtime, pulling exactly the context they need. Humans put knowledge in. Agents navigate the structure.
Structuring knowledge uses LLM calls — that’s the investment. Once structured, everything is pre-indexed and retrievable via MCP at the cost of a database query.
Any MCP-compatible client — Cursor, Claude Code, custom agents — reads your project’s structured knowledge with no additional LLM cost on top of whatever you’re already running.
built for
Think with AI, capture decisions as they land. Start each new session from accumulated structured knowledge — not a blank context window.
QA agents push feedback. Leadership adds directives. Engineers record progress. Every agent reads from the same governed source of truth — via MCP.
Three months in, a new contributor joins. They query the knowledge base via MCP and get the full decision history — what was decided, what was tried, what was rejected — in seconds.
Conversations in. Decisions, constraints, and direction out — organized, connected, and always current.
MCP-native · self-organizing · zero read cost
join the waitlistCurrently in closed beta. Sign up to reserve your spot.