Hippocortex/Docs
GitHubDashboard

Knowledge Graph

The Knowledge Graph extracts entities and relationships from your agent's memories, creating a connected web of knowledge that enhances retrieval.

Feature flag: HIPPOCORTEX_GRAPH_PACKAGE (enabled by default)

Entity Extraction

Two methods:

  • Regex baseline (@hippocortex/graph) — fast, zero-latency, runs on every event
  • LLM NER (GPT-4o-mini) — higher accuracy, runs during distillation, bilingual DE/EN. Uses LRU cache (1K entries, 10min TTL) and rate limiter (100 calls/min). Feature flag: HIPPOCORTEX_LLM_NER

Entity Types

person, organization, technology, product, service, location, concept

Graph Structure

Nodes: entity name, type, metadata, evidence references. Edges: requires, improves, contradicts, supersedes, related_to with weights (0-1).

Graph-Enhanced Synthesis

When HIPPOCORTEX_GRAPH_SYNTHESIS is enabled, synthesize combines vector retrieval with 1-hop graph traversal for richer context.

API Endpoints

MethodEndpointDescription
GET/v1/graph/fullFull graph (nodes + edges)
GET/v1/graph/nodesList graph nodes