Skip to content

Knowledge Service

Cognition Knowledge Service is the typed knowledge surface for agent reasoning — pages with first-class relations, lexical and hybrid retrieval, ingest lifecycle. Distinct from runtime knowledge banks; this is the standalone authority's knowledge surface.

What Knowledge Service Owns

ConcernSurface
Typed page lifecycleSave / list / load / delete
Lexical retrievalKeyword / phrase search
Hybrid retrievalLexical + vector
First-class relationsTyped relation graph between pages
Ingest lifecyclequeued → running → completed/failed
Page metadataService-derived

Pages are typed. The relation graph is first-class — you can query "what is connected to this page" and get a typed graph result, not a flat list.

First-Class Relations

PropertyValue
StorageTyped relation graph
CardinalityPer-relation typed
Cross-scopeForbidden (cognition scope-bound)
MutationThrough admitted relation contracts
Query"What is connected to X" returns typed graph

This is what distinguishes Cognition Knowledge from a flat search-index. A reader (or agent) can navigate the relation graph; reasoning can traverse "what derives from what."

Ingest Lifecycle

StateMeaning
queuedAwaiting ingest
runningIngest in progress
completedIngest finished successfully
failedIngest failed; reason recorded

Interrupted local tasks become explicit failed-state evidence on reopen. The platform does not silently lose ingest progress; on reopen, the failure is visible.

An agent needs to recall information about a topic during a turn.

  1. Lexical search. Agent issues a query under KnowledgeService.Search.
  2. Service returns matches. Pages matching the query, with service-derived relevance metadata.
  3. Agent uses results. Brain layer composes the response with the recalled knowledge.

The agent did not have to scan all pages itself; the service provides the typed search surface.

Reader Scenario: An Agent Navigates The Relation Graph

An agent has a page about "interview tomorrow." Related pages might include "user's preferred talking points," "user's career goals," "user's anxiety triggers."

  1. Navigate from anchor page. Agent invokes KnowledgeService.RelatedPages(pageId).
  2. Service returns typed graph. Connected pages with typed relation kinds.
  3. Agent traverses. Reasoning includes the related context.
  4. Composed response. Reflects the agent's broader understanding.

First-class relations enable structural reasoning. A flat search-index cannot answer "what's connected to this."

Reader Scenario: An Ingest Pipeline Fails Mid-Run

A user ingests a large document set; halfway through, the process is interrupted.

  1. Ingest started. State queued → running.
  2. Process killed. External interruption.
  3. On reopen. Ingest state is detected as failed (or running with stale heartbeat that times out to failed).
  4. Failure visible. User sees typed failure with reason — not silent partial completion.
  5. Resume / restart. User can resume under admitted resume contract or restart the ingest.

The platform makes interruption visible. Silent partial completion would let stale state masquerade as completion.

Cleanup Of Knowledge

Knowledge pages, like memory records, are eligible for digest cleanup.

Cleanup driverWhat it cleans
DigestPages with broken refs / stale support
Explicit deleteUser-driven destructive delete
ArchiveReversible archive (digest first pass)

Cleanup is traceable to refgraph reasoning. A page archived because of broken refs has the broken refs as the explanation.

Cognition Scope For Knowledge

Knowledge pages live within a Cognition scope. Cross-scope references are forbidden. This means knowledge from one agent's scope does not leak into another's.

Boundary Summary

ConcernOwner
Page storageCognition Knowledge Service
Relation graphCognition (first-class)
SearchLexical / hybrid retrieval
IngestTyped lifecycle
CleanupDigest + explicit delete

What Knowledge Service Does Not Do

ConcernWhy not
Mutate kernelsKernels are core truth; advisory cannot demote
Cross-scope referencesHard reject
Free-form unstructured ingestPages are typed
Implicit relevanceService-derived metadata is explicit

Source Basis

Nimi AI open world platform documentation.