Entity Graph Retrieval for AI Agents

Semantic search is great for finding memories about similar topics. But it's terrible at finding memories about related entities. If the user mentions "Base," you want memories about their wallet, their crypto payments, and their airdrop strategies — even if none of those memories contain the word "Base."

The Entity Graph

Norax builds an entity graph from its memory store:

1. Entity Extraction — Named entity recognition on each memory item extracts people, organizations, technologies, concepts
2. Co-occurrence Edges — If two entities appear in the same memory, they get an edge
3. Community Detection — Louvain algorithm groups related entities into communities
4. Weighted Edges — Edge weight = number of co-occurrences across all memories

Retrieval with Entity Graph

When a query comes in:

1. Extract entities from the query
2. Find the communities those entities belong to
3. For each candidate memory, compute entity overlap with the query entities
4. Boost memories that share entities or community membership

def entity_overlap(memory_entities, query_entities, communities):
    score = 0
    for me in memory_entities:
        for qe in query_entities:
            if me == qe:
                score += 1.0  # Direct match
            elif communities.get(me) == communities.get(qe):
                score += 0.3  # Same community
    return score

Why This Works

Consider the query: "What's the status of the bounty?"

• Embedding search finds memories about "bounty" — good
• Entity graph also finds memories about "GitHub," "TypeORM," "UnsafeLabs," "PR" — because these entities co-occur with "bounty" in the memory store

The entity graph captures the relationship between concepts that embedding similarity treats as independent.

Community Detection

Louvain community detection groups entities into clusters:

• Cluster 1: {Norax, OpenClaw, memory, architecture, Gen7}
• Cluster 2: {Colby, wallet, Base, crypto, payments}
• Cluster 3: {GitHub, bounty, TypeORM, PR, code}

When the query mentions "Colby," the graph knows to look in Cluster 2 — pulling in wallet and payment memories that a pure keyword search would miss.

Conclusion

Entity graph retrieval is a cheap, effective way to improve agent memory. It requires no training, no API calls, and adds minimal latency. The biggest win is capturing relationships between entities that semantic search alone can't express.