Chat logs, wikis, and vector stores were built for humans reading. Pragmus is built for agents acting. It holds the working state of what's known, claimed, decided, and asked — current, contradiction-aware, and ready to load.
A coding agent decides on a token strategy. A research agent learns a constraint. A planning agent commits to a sequence. Today, those decisions land in a session transcript, get auto-summarized, and disappear. The next agent starts from zero — or worse, from a snapshot of yesterday's assumptions that are no longer true.
The shape they need isn't a document. It isn't a graph database. It's a substrate — small, structured, queryable, and self-indexing — that holds the state-of-the-world an agent has to read before it can act competently.
Pragmus is that substrate. how it works →
Your own working memory, shared with the agents you run. Your next session loads what you already figured out.
solo →Drift gets surfaced, not buried. Decisions don't live in DMs. The team's record stays current.
teams →Pragmus is the context endpoint your agents write to and read from. One MCP server, every agent on the same record.
agents →Pragmus ships with a dashboard chat — an in-browser agent that captures items and answers questions over the substrate using the same MCP tools an external client would. You describe what happened; it drafts the item, you confirm. You ask the state of something; it queries the right tools and replies. Useful when you're working from a browser and haven't wired up a client yet.
No SDK. No per-agent adapter. Same tools, same writes, same reads, regardless of which agent runtime is on the other side.
Works with any MCP-compatible client. Documented setups: Claude Code, Claude Desktop, Cursor. Anything that speaks MCP can connect.