The Librarian: auto-linking your knowledge graph

A personal knowledge base becomes exponentially more useful when memories are interconnected. Manually curating hyperlinks between hundreds of documents quickly becomes a chore — so an AI memory system needs an autonomous background mechanism to discover and weave connections between related memories.

The core challenge isn’t just calculating similarity: it is keeping this relationship discovery fast, decoupled, and completely isolated from real-time ingestion.

The Mechanics of Autonomous Graph Discovery

Rather than forcing heavy all-pairs calculations during synchronous user ingestion, relation discovery operates as an event-driven background pipeline:

1. Decoupled Event Stream: Ingesting a note returns immediately to ensure sub-millisecond response times. A background worker picks up the newly indexed record.
2. Feature Extraction: High-salience key terms and structural entities are vectorized.
3. Similarity Gating: Candidates in the same workspace are scored against a confidence threshold, ensuring only genuine relationships form edges in the graph.

High-Level Algorithm

function DiscoverGraphRelations(newDocument, candidatePool):
    # Step 1: Extract feature representation
    targetVector = ExtractFeatureWeights(newDocument)
    
    discoveredEdges = []
    
    # Step 2: Evaluate similarity against candidate pool
    for candidate in candidatePool:
        if candidate.id == newDocument.id:
            continue
            
        score = CalculateCosineSimilarity(targetVector, candidate.vector)
        
        # Step 3: Threshold gate ensures quality over quantity
        if score >= RELATION_CONFIDENCE_THRESHOLD:
            discoveredEdges.append({
                sourceId: newDocument.id,
                targetId: candidate.id,
                weight: score,
                relationType: "RELATED_TO"
            })
            
    # Step 4: Persist bi-directional graph edges
    GraphStore.insertEdges(discoveredEdges)

Architectural Benefits

• Non-Blocking Write Path: Document ingestion never waits on graph computation.
• Dynamic Exploration: Agents can traverse 1-hop and 2-hop graph neighborhoods over the Model Context Protocol (MCP) to gather comprehensive context during complex research tasks.