How Pragmus organizes what your agents produce

Four kinds of knowledge. Four intents.

Every entry declares what kind of thing it is. The system treats each one differently — because an observation isn't a decision, and a question isn't a claim.

observe

A factual event or data point. Ephemeral; ages out at milestones.

claim

An assertion open to validation. Contradicted by a later inverse claim.

decide

A binding decision or task. Closed by execution; invalidated if reversed.

inquire

An open question. Closed when answered; invalidated if the premise dies.

The vocabulary is small on purpose. An agent (or its operator) doesn't pick a folder — it picks a stance. The system handles the rest.

You write items. Pragmus indexes the aspects.

Each item's body gets sliced into claim units — the atomic things the system reasons about. A unit is the smallest piece of memory that still makes sense on its own.
The LLM extracts threads per unit (concept labels — e.g. auth-flow, migration-v2) using your project's subjects as vocabulary guidance.
Inline #tags in the prose declare narrative continuity that the LLM might miss.
Topics are how you curate the substrate. You author a topic with a prompt (or a description); the system threads it and pulls every unit whose threads or tags overlap. A cluster proposer surfaces candidate topics from emergent thread communities — you accept the ones that matter, the rest never materialize.
Topics are perspectives, not folders: a unit lives in every topic whose anchors it carries. Filter further with scope — intent, subject, or recency — layered on top of thread overlap.

No taxonomy to design upfront. The system suggests structure; you decide what's real.

State-check, with a human in the loop.

When a new claim closes, contradicts, or invalidates an older unit, the system spots it and proposes a resolution. The human (or the agent's operator) confirms or dismisses. Nothing changes the record without approval.

That's the trade Pragmus makes: the system does the boring work of noticing drift; you stay in charge of what counts as resolved.

Milestones let you say "the next era starts here." Old items remain searchable; they just stop anchoring the active topic graph. The result is a working context that fits in an agent's window without losing the archive.

Sandboxes let you isolate a parallel hypothesis — a planning cycle, an experiment, a separate client — without crossing wires with the main graph.

Zero LLM cost on reads. Just database queries.

knowledge, topic, area, open, search, walk, threads, thread, item, unit — every read surface is a structured query. No re-summarization, no re-clustering at read time. Your agent connects via MCP and navigates pre-organized state.

Writes are where the LLM work happens (extraction, threads, summaries, state-check proposals). Reads are free.

One picture.

Four kinds of knowledge. Four intents.

You write items. Pragmus indexes the aspects.

State-check, with a human in the loop.

Zero LLM cost on reads. Just database queries.

One picture.

One substrate. Three lanes of use.