Hybrid retrieval
Full-text and vector retrieval are fused and reranked, while graph relationships support multi-hop context.
← All products
Active research
Persistent memory for AI agents
Context Heavy
A multi-tenant knowledge graph and hybrid retrieval system that stores agent context across sessions.
Verified facts
- 71.2% recall@10 on a 304-question LoCoMo slice
- Hybrid vector, text, graph, and reranking
- Go + PostgreSQL 16 + pgvector
- Benchmarked with the public CH-Bench harness
System brief
What it does.
Context Heavy stores entities, documents, temporal updates, and typed relationships in PostgreSQL with pgvector. Agents can retrieve through REST and MCP instead of rebuilding context from flat prompt history.
Temporal memory
Supersession and recency let the system distinguish current facts from stale ones.
Workspace isolation
Nodes, edges, credentials, and queries are scoped to a tenant workspace.
Measured, with caveats
On a two-conversation, 304-question LoCoMo retrieval slice it led the compared systems on ranking metrics, with a slower 1.7-second p50.
Evaluation
Published benchmark result.
On a retrieval-only LoCoMo slice containing two conversations and 304 questions, Context Heavy reached 71.2% recall@10 and 60.1% nDCG@10. In that run, supermemory reached 63.3% recall and 45.4% nDCG; gbrain reached 58.0% and 38.2%; mem0 reached 20.6% and 17.4%.
Context Heavy’s 1.7-second p50 was slower than supermemory’s 0.7 seconds. The remote cross-encoder reranker increased latency. The slice is smaller than full LoCoMo and is not comparable with answer-judge scores. CH-Bench publishes the adapters, dataset notes, and reproduction commands.
Inspect CH-Bench ↗