You think out loud across sessions. You decide, change your mind, hit dead ends, revisit. Pragmus is where that record lives — organized as you go, readable by you and by every agent you run.
// you write what just happened — → ingest decide "ship migration fix behind a flag #migration-v2" → ingest claim "deadlock is in migration ordering, not the lock" → ingest inquire "does the rollback path drop the index?" ▼ state-check proposes (closed · invalidated · contradicted) ▼ you confirm or dismiss // next session, the agent reads — → knowledge Migration v2 (topic, 12 items, #migration-v2) Decision — feature-flag new ordering, ship #migration-v2 Recent claim — staging deadlock reproduced after 3rd reboot Open — does the rollback drop the index? ► 3 open · 23 items · 0 LLM calls · ready to act
Every new session starts cold. Yesterday's decision isn't loaded; last week's constraint isn't surfaced; the dead-end you walked down gets proposed again. Notes are unstructured prose. Wikis hold still while your thinking doesn't. Vector stores retrieve what's nearby, not what's current. You need a working record that updates as you do, with structure you don't have to maintain.
From any MCP-compatible client — Claude Code, Claude Desktop, Cursor — your agent calls ingest when something's worth keeping. Or use the dashboard chat: describe what happened in plain prose, the chat drafts an item with intent + subjects + tag suggestion, you confirm before it lands.
The body is sliced into claim units — atomic, one claim each. The LLM coins concept labels (threads) per unit while it's already reading. Two items about auth-flow are connected by that exact label, looked up in one query — no fuzzy match. Drop a #tag when you want explicit narrative continuity that future-you (or any agent) can walk.
A topic is a saved view you author with a prompt: "the migration v2 work", "open questions on auth". The system threads the prompt and pulls every unit whose threads or tags overlap. Next session, you open the topic and the narrative, the bulleted findings, and the open questions are loaded — current as of right now.
Filter further with scope — intent, subject, recency — layered on top. The cluster proposer surfaces candidate topics from emergent thread communities; accept the ones that matter. Topics aren't folders; a unit lives in every topic whose anchors it carries.
When a new claim closes, contradicts, or invalidates an older unit, Pragmus proposes the resolution. Claims get contradicted; decisions and inquires get closed (positive) or invalidated (negative). Per-unit, not per-item — an item can have some claims contradicted and others still standing.
Proposals land in the review queue with the evidence that triggered them. Confirm to apply; dismiss to leave the older item live. The system spots; you decide.
When you finish a phase — a launch, a sprint, a research push — drop a milestone. Done and ephemeral items archive in bulk; the snapshot of your topic tree is preserved for later browsing. Your active context stays small.
Spinning up a parallel hypothesis — a new architecture, a different positioning, a what-if? Create a sandbox. Items in it cluster only with each other. When you decide which version wins, you keep what's worth keeping and the rest doesn't bleed into your main project.
Works with any MCP-compatible client. Documented setups: Claude Code, Claude Desktop, Cursor.