Graph Knowledge & Tri-Search: Engram’s Unique Differentiator

Overview

Engram’s Graph Knowledge (Gk) is a unique differentiator in the AI Periodic Table framework. It represents the critical third layer of Engram’s emergent tri-search capability, enabling relationship-based reasoning and multi-hop traversal.

Tri-Search Architecture

Engram’s memory retrieval uses three complementary search layers combined via Reciprocal Rank Fusion (RRF):

  1. Keyword Search (BM25/Lexical): Exact phrase matching, acronym lookup, terminology search
    • Storage: Zep sessions with chunked messages
    • Enables: Exact term matching, acronym resolution, technical term lookup
  2. Vector Search (Semantic): Conceptual similarity, paraphrase matching, meaning-based retrieval
    • Storage: memory_embeddings table (pgvector)
    • Embedding Model: text-embedding-3-small (1536 dimensions)
    • Enables: Conceptual similarity, paraphrase detection, context-aware retrieval
  3. Graph Search (Knowledge Graph): Relationship traversal, entity connections, multi-hop reasoning
    • Storage: Zep Temporal Knowledge Graph
    • Enables: Entity relationship discovery, multi-hop reasoning, temporal fact tracking, cross-session knowledge assembly

Graph Knowledge Structure

Node Types

  • Fact (Cyan #00d4ff): Semantic facts extracted from conversations
  • Entity (Purple #a855f7): People, projects, concepts
  • Memory (Green #10b981): Episodic memory (conversation sessions)
  • Topic (Amber #f59e0b): Conversation topics/themes
  • Metadata (Gray #6b7280): Source, filename, tenant tags

Edge Types

  • concerns: Episode relates to topic
  • source: Fact comes from source
  • related_to: Entities are related
  • mentions: Content mentions entity

Knowledge Graph Interface

URL: https://engram.work/memory/graph

Features:

  • Interactive force-directed graph visualization
  • Search and filter by query, node type, and degree
  • Statistics dashboard (total nodes, edges, average degree, node type breakdown)
  • Node details panel (content, metadata, connections)
  • Tri-search context explanation

API Endpoint 

GET /api/memory/graph?query=<search_term>

Returns graph nodes, edges, and statistics including:

  • Total nodes and edges
  • Node types breakdown
  • Average and maximum degree
  • Filtered by query if provided

Use Cases

  1. Entity Discovery: Find who worked on a project by traversing graph relationships
  2. Project Timeline: View chronological project history via episode traversal
  3. Knowledge Gap Analysis: Identify isolated nodes (low degree) needing more context
  4. Multi-Hop Reasoning: Answer complex queries requiring multiple relationship hops

How It Works

Graph Knowledge is built automatically from:

  • Conversations: Entities, facts, and topics extracted from agent conversations
  • Documents: Entities and facts from ingested documents
  • Episodes: Session summaries become memory nodes with topic links

Users can also manually add facts via POST /api/memory/facts

Observability

Available:

  • Graph structure visualization
  • Node details (content, metadata, connections)
  • Statistics (total nodes, edges, degree metrics)
  • Search transparency (which nodes match query)

Planned:

  • Tri-search breakdown (which layer contributed each result)
  • Retrieval path visualization
  • Graph analytics (centrality, community detection)
  • Time slider for temporal graph views

Technical Implementation

  • Backend: Zep Cloud Temporal Knowledge Graph, entity extraction, fact storage
  • Frontend: react-force-graph-2d visualization, interactive filtering
  • Data Flow: User query → GET /api/memory/graph → _build_graph() → Zep APIs → Graph + Stats → Visualization
  • Full guide: docs/features/memory/graph-knowledge-tri-search.md
  • AI Periodic Table: docs/wiki/ai-periodic-table.md
  • Roadmap: docs/ai-periodic-table-roadmap.md