AI Agent Knowledge Graph

Use text or vector search for direct matches. Use an AI agent knowledge graph for typed relationships: who worked on a project, which tool it uses, and which grounded observations belong to the same entity.

A reliable source-backed AI knowledge base keeps that graph beside inspectable source memories and maintained pages. Graph context supplies nearby facts; it does not become authority or replace provenance and review.

An entity is a named thing such as a person, project, tool, repo, or concept. A relation connects two entities. An observation is a grounded statement attached to an entity.

The goal is practical retrieval. The graph should help the next session find the right surrounding context, not become a hand-maintained ontology.

Post-ingest enrichment can link entities, deduplicate overlapping captures, enrich titles, grow matching pages, and update effective confidence based on memory type, access, and age. Optional local models or API keys can make extraction and graph linking richer.

Claude Code skills can also help because the agent already has language judgment during capture and distillation.

Imported Markdown vaults can carry useful links. Wenlan can preserve those relationships as graph signal while still treating imported content as context that should prove itself useful in recall.

Wikilinks are helpful, but they are not the same as source-backed page provenance. A page claim is stronger when it traces to specific source memory IDs.

Hybrid retrieval can combine vectors, full-text search, pages, and graph context. The graph adds nearby related facts when text similarity alone would miss why a memory matters.

Treat graph context as support, not authority. The source memory, page provenance, and local git history remain the inspectable record.

Graph data can be wrong when a name is ambiguous, a project changes direction, or an imported link was noisy. Use review, corrections, spaces, and forget when stored context needs cleanup.

If graph context feels like it is mixing unrelated worlds, check the active space before changing the data model. Many apparent graph problems are actually context-boundary problems.