Knowledge

How the agent accumulates experience — the knowledge lifecycle, the area map, dossiers, contradictions, bitemporality.

Knowledge is the verified layer of experience: statements about the environment, tools and solutions, extracted from real trajectories. Unlike a "memory dump", every piece of knowledge carries an author, provenance, a status and a change history. The Knowledge page (/knowledge) serves both humans and agents (the knowledge_explore tool).

Knowledge is about the project and its environment; personal preferences and people's roles live in a separate layer of facts — see Personalization.

#Lifecycle

proposed → confirmed → (challenged → corrected | superseded | invalidated)

• Proposed — extracted from a run, not yet verified.
• Confirmed — applied and worked (or confirmed by a human).
• Challenged — a contradiction was found; under review.
• Corrected / Superseded / Invalidated — terminal outcomes with the history preserved: who, when and why.

"Alive" knowledge is what the system is ready to hint to the agent. Stale ("at risk") items are shown with a marker and re-evaluated.

#The area map

A treemap groups knowledge by topics and entities: tile size — the amount of knowledge, color — the dominant state, a red badge — contradictions. Clicking an area applies a sticky filter to the knowledge list.

The Tidy up button runs the judge model (the same one the quality loop uses): it merges duplicates and variants ("playwright", "playwright e2e") into canonical areas and hides garbage (file names and the like). Knowledge records are not modified — only the display "lens" is.

#Search and filters

• Compact filters: state, visibility scope, type (claim/observation), origin, topic, entity.
• Semantic search — a question in plain words; results explain why they matched.
• Bitemporal filters — look at the base "as it was": at an event date or as of a moment of knowledge (who knew what when). Essential for incident retrospectives.

#The knowledge dossier

Clicking a knowledge item opens its dossier: the proposition, status and transition history, primary sources (links to specific run events), relations to other knowledge (supports/contradicts/derived from), usage statistics.

From the dossier: challenge, correct, invalidate — with a reason and evidence. You can also challenge or invalidate straight from the run trace — the ⚑ button on the event that references the knowledge: the doubt is captured right where you noticed it. These are human actions; the agent flags doubts, but you move the lifecycle.

#Contradictions

The Contradictions tab collects pairs of knowledge items that state different things about the same subject: a side-by-side comparison and a resolution (keep one, correct, both are stale).

#Research sessions

The current question, filters and selected area are saved as a session you can return to later. The agent explores the base too, through the knowledge_explore tool (orient/search/expand/trace/compare/synthesize), and its research runs appear in a separate tab.

The mechanics in depth — docs/knowledge-model.md, docs/knowledge-evolution.md, docs/human-knowledge-explorer.md.