Skip to content

Knowledge Platform — overview

Target

This page describes intended behaviour or structure. No implementation is claimed unless a section is explicitly marked Current implementation or Evidence.

Rewritten 2026-10-09 (design plan P4)

This page replaces the June-2026 overview, which listed registries of patterns and prompts, replay and runtime memory as services. Those are later-stage rows now; the 2026 modules are ingestion, retrieval and indexes, context assembly, artifacts and lineage, promoted memory and repository intelligence. The old page is archived under catalogs 2026-06.

Knowledge is what the organization knows; the graph is how things relate; memory is what happened and what was learned; a skill is how something should be done. The Knowledge Platform keeps the first three as durable, governed, provenance-carrying assets that outlive any model, and it builds the context every attempt runs on.

Platform PF-04 Knowledge Platform (Knowledge & Memory)
Bounded contexts BC-09, BC-10, BC-11, BC-12, BC-13, BC-39
Services 24 (1 hosts, 17 modules, 0 external, 6 later) — 7 modelled
Gateways —
Status target

Purpose

  • Source of truth stays the source. Immutable snapshots of real sources with ACL propagation; nothing is retrievable before its permissions are known.
  • Scope before ranking. Tenant, project and ACL filters run in the backend before top-k (ADR-0013); citations carry source version, chunk and index revision.
  • History is not overwritten. Graph facts are temporal and carry provenance and a trust tier (source fact, inferred fact, agent observation, unverified memory).
  • Context is engineered, not dumped. A planner budgets sections, ranks, deduplicates and compresses; the snapshot says what was used and why.
  • Memory is governed. Observations become memory only through classification, evidence and review; agent opinions never become facts on their own; reinforced procedures become skill proposals.
  • Repositories are indexed, not pasted. Symbols, dependencies, calls, ownership and tests per revision, served as read tools.

Bounded contexts and services

Context Owns Services (granularity)
BC-10 Sources & Ingestion Sources, collections, snapshots, pipeline, deletion lineage KnowledgeIngestionService (module); code, documentation, library and template knowledge rows merged as extractors
BC-11 Retrieval & Indexes Index revisions, graph facts, retrieval queries VectorMemoryService (module, historical name); search, embedding and metadata-index rows merged; KnowledgeGraphService later
BC-09 Context Assembly Context snapshots and items ContextBuilderService (module)
BC-12 Artifacts & Lineage Artifacts, version chains, lineage ArtifactService (module); lineage, versioning, preview and artifact-memory rows merged
BC-13 Promoted Memory Memory items, review decisions, consent MemoryPromotionService (module, added); prompt, runtime and decision memory rows later
BC-39 Repository Intelligence Repository index sets, symbols, dependencies, ownership RepositoryIntelligenceService (module)
Host Composition, vector profile, embedding client, connectors KnowledgeHost (standalone)
Service Context Granularity Status
ArtifactMemoryService BC-12 module of host unmodelled
ArtifactPreviewService BC-12 module of host unmodelled
ArtifactService BC-12 module of host target
CodebaseKnowledgeService BC-10 module of host unmodelled
ContextBuilderService BC-09 module of host target
DecisionMemoryService BC-13 later stage unmodelled
DocumentationKnowledgeService BC-10 module of host unmodelled
EmbeddingService BC-10 module of host unmodelled
KnowledgeGraphService BC-11 later stage unmodelled
MemoryPromotionService BC-13 module of host target
KnowledgeHost BC-11 standalone host target
KnowledgeIngestionService BC-10 module of host target
KnowledgeReplayService BC-11 later stage unmodelled
KnowledgeSearchService BC-11 module of host unmodelled
LibraryKnowledgeService BC-10 module of host unmodelled
LineageService BC-12 module of host unmodelled
MetadataIndexService BC-11 module of host unmodelled
PatternCatalogService BC-13 later stage unmodelled
PromptMemoryService BC-13 later stage unmodelled
RepositoryIntelligenceService BC-39 module of host target
RuntimeMemoryService BC-13 later stage unmodelled
TemplateKnowledgeService BC-10 module of host unmodelled
VectorMemoryService BC-11 module of host target
VersioningService BC-12 module of host unmodelled

Role in the Factory

flowchart LR
    Src["Sources<br/>Git, Azure DevOps, SharePoint, Confluence, documents, SQL, APIs, tickets, MCP resources"] --> KP["Knowledge Platform<br/>KnowledgeHost"]
    KP -- "context snapshots, tools, artifacts" --> Mesh["Agent Mesh"]
    Mesh -- "observations, outputs, checkpoints" --> KP
    KP -- "Embed, Rerank" --> MC["Model & Compute"]
    KP -- "promotion approvals" --> CP["Control Plane"]
    KP -- "knowledge and repository tools" --> Tools["Tools & MCP"]
    KP -- "decisions, redaction, credentials" --> Gov["Governance"]
    KP -- "evidence, snapshots" --> Obs["Observability"]
Hold "Alt" / "Option" to enable pan & zoom

The flows (original handoff §68, §69, §19)

flowchart TB
    subgraph Ingest["Ingestion"]
        A["Snapshot"] --> B["Parse, normalize"] --> C["Metadata, ACL"] --> D["Chunk"] --> E["Text index · Vector index · Graph extraction"] --> F["Provenance"] --> G["Published"]
    end
    subgraph Learn["Learning loop"]
        R["Run completed"] --> O["Observations"] --> Cl["Classify"] --> Ev["Evidence and policy"] --> Pr["Promoted memory or skill proposal"]
    end
    subgraph Ctx["Context assembly"]
        T["Task"] --> Pl["Plan sections"] --> Rt["Retrieve, expand, memory, repository state"] --> Rk["Rank, dedupe, compress, budget"] --> Sn["Snapshot"]
    end
    G --> Rt
    Pr --> Rt
Hold "Alt" / "Option" to enable pan & zoom

Component view

flowchart LR
    subgraph P["Knowledge Platform"]
        KnowledgeHost["KnowledgeHost"]
    end
    P --> PolicyEngineService["PolicyEngineService"]
    P --> CostUsageService["CostUsageService"]
    P --> AuditService["AuditService"]
    P --> ModelRouterService["ModelRouterService"]
    P --> RunControlService["RunControlService"]
    P --> AgentRegistryService["AgentRegistryService"]
    P --> ModelGatewayHost["ModelGatewayHost"]
    P --> ControlPlaneApi["ControlPlaneApi"]
    P --> GovernanceHost["GovernanceHost"]
    P --> IntegrationHost["IntegrationHost"]
    P --> ext_identity_platform["ext-identity-platform"]
    P --> IntegrationService["IntegrationService"]
    P --> SecretBrokerService["SecretBrokerService"]
    P --> ProjectService["ProjectService"]
Hold "Alt" / "Option" to enable pan & zoom

Built from the ConnectSoft estate

Component Use here Status
ConnectSoft.Extensions.DataIngestion pipeline primitives (chunking, embedding, upsert) extend
ConnectSoft.Extensions.DataStorage.* artifact content port exists
ConnectSoft.Extensions.AI (embeddings, SK vector connectors) vector profiles per ADR-0013 extend
ConnectSoft.Extensions.Compliance classification and redaction exists
Microservice template, MessagingModel Orleans adapter host, outbox and inbox exists
ConnectSoft.MediaIngestion transcript connector candidate extend
Ingestion lifecycle, retrieval scope contract, graph facts, context snapshots, artifacts registry, memory lifecycle, repository indexes the platform itself develop

Modernization (2 to 3 year horizon)

  • Agent memory standards and governed learning (D32, D33): memory classes with lifecycle, review and consent instead of free-form memory stores.
  • Provenance and content trust tiers feed guardrails: retrieved content is data with a tier, never instruction.
  • Repository intelligence as MCP tools (ADR-0022); AGENTS.md and skill-package conventions become ingested sources.
  • Temporal knowledge graphs per domain with a federated query layer.

Current implementation

Chunking, embeddings and top-k search exist in the agent host sample with post-search filtering and only the in-memory connector; ingestion primitives exist without lifecycle; no graph, artifact registry, context snapshot, memory lifecycle or repository index exists. See the plane index.

Final-state summary

One host, six modules, three stores; every retrievable item has a scope, a version and a trust tier; every context is explainable; every artifact has lineage; every memory was reviewed or is labelled as an observation.