Skip to main content
Forge Knowledge stores durable information that people and agents can retrieve and cite. Artifacts can be authored in Forge, synchronized from an external source, recorded by an agent, or promoted from eligible memory. Use the Knowledge console to find and maintain artifacts. The lifecycle below connects sources, access, retrieval, evidence, and review.

The system at a glance

The Artifacts and lifecycle section below covers these boundaries in detail.

Knowledge is not agent memory

Knowledge artifacts are durable entries in a governed scope. Other readers and agents can use them only when their own access reaches that scope and the source allows them. Agent memory is context retained for a particular agent: session and channel summaries, long-term memories, facts, and episodes. It has separate permissions and operator controls. Memory does not become shared knowledge merely because an agent recalls it. An eligible long-term memory may later be promoted into a provisional knowledge artifact; after promotion, that artifact follows the normal Knowledge access, retrieval, evidence, and review lifecycle. See Use Knowledge in agents for the practical boundary.

Artifacts and lifecycle

A knowledge artifact is one durable entry that Forge can browse, retrieve, cite, and govern. Artifacts include written knowledge such as lessons and policies, structured records, files and documents, synchronized items, and memories that have been promoted into Knowledge. The artifact’s kind, origin, grade, and access boundary are independent. Read them together instead of treating one label as a complete trust decision.

Kind and origin describe different things

An authored kind describes the artifact’s purpose: Lesson, Policy, How-to, Reference, FAQ, or Note. Older or system-produced artifacts can be unclassified. Structured records use their registered record kind instead. Origin describes how the artifact entered Knowledge:
  • Authored in Forge: a person or API client maintains a Forge-owned artifact or typed record.
  • Recorded by an agent: an agent deliberately writes a durable case during its work.
  • Synchronized: an external system remains the source of truth; synchronization maintains the stored item and its source access.
  • Promoted from memory: an eligible long-term agent memory becomes a provisional artifact in an allowed Knowledge scope. Automatic Personal Agent promotion stays in its owner’s personal scope, rather than adopting the project or run that recalled the memory.
An origin is provenance, not authority. A synchronized artifact can be restricted to selected people, and an agent-authored artifact does not bypass the target scope’s write or read rules.

Scope and source access decide visibility

Every artifact belongs to a knowledge scope. Shared organizational scopes include company, department, team, factory, and project; agents can also have their own scope, and personal knowledge remains personal. The reader’s current knowledge:read closure is the first boundary. A synchronized source can narrow that boundary further. If the source item is restricted, Forge shows it only to mapped people and agents acting on behalf of an allowed person. A broad role or platform administration does not turn a source-restricted item into open knowledge. Collections, search, citations, history, and review all retain these checks. Scope and source access do not change when an artifact’s grade changes. Likewise, putting artifacts in a collection organizes them but does not grant access.

Grade, outcomes, and verification are separate

Cases can be: Search combines relevance with the configured weight for the current grade. A proven grade does not assert that a person verified the current text, that every use succeeded, or that the source is unrestricted. When work cites a case, Forge can associate outcome evidence with that use. Signals can come from domain outcomes, operator ratings, Judge results, failed runs, or feedback on a chat answer. The effective signal contributes to the case’s success and failure history. Automatic transitions are deliberately limited; human review handles decisions such as restoring, retiring, or resolving a conflict. Company and global cases need human approval before outcome learning makes them proven. Verification records that a person reviewed the current text. It has its own steward and due date and does not by itself change the grade or search order. Editing the text clears the previous verification; the new text must be reviewed again.

Human review closes the loop

Review items are specific requests, not a single generic approval queue. Depending on the reason, a steward may:
  • approve an earned promotion or leave the case provisional;
  • resolve contested evidence by restoring, resetting, or retiring a case;
  • choose between conflicting cases;
  • merge a duplicate pair when it still satisfies the merge rules;
  • retire or keep stale knowledge;
  • verify current text when verification is due; or
  • accept or reject an explicit request for review.
The actions offered in Console depend on the review kind and current state. Open the artifact’s source, evidence, uses, versions, and grade before deciding. “Verified” and “proven” answer different questions.

Operate the lifecycle in Console

  1. Open Knowledge to search or browse the artifacts you can read.
  2. Open an artifact to inspect its overview, source, evidence, uses, and history separately.
  3. Use Sources for synchronization state and source-owned content. Removing a source removes what it synchronized; retiring one artifact is different.
  4. Use Review for steward decisions and the activity trail.
  5. Use Health when an artifact is stored but not searchable or a source, embedding, or integrity check needs attention.
For the complete synchronized-source contract and Forge-owned structured facts, see Synchronized knowledge sources and its Authored records section.

Choose the right entry path