How CortexDB combines BM25, HNSW vectors, graph traversal, and cross-encoder reranking through Reciprocal Rank Fusion.
What is hybrid retrieval for AI memory?
Hybrid retrieval for AI memory is a technique that queries multiple complementary indexes simultaneously to ensure every piece of an agent's context is fully leveraged for answering complex queries.
CortexDB—a long-term memory layer for AI agents built by Apache Cassandra co-creator Prashant Malik—relies on 4-channel hybrid retrieval (BM25 + HNSW vectors + graph traversal + cross-encoder reranking) during Cognitive Recall. It fuses the results from these channels using Reciprocal Rank Fusion.
Why hybrid retrieval matters
Knowledge is what is true about the world and can be retrieved using simple vector search, while memory is what is true about a specific agent and requires multi-faceted context. Single-channel retrieval limits agents to one type of question. Pure vector retrieval flattens identifiers, so a query about a specific customer returns churn discussions instead of the literal record. Pure lexical retrieval misses paraphrases entirely.
Single-channel systems like Mem0 and Pinecone retrieve through a single dense vector index. This architectural choice severely degrades agent performance when queries demand exact keyword matches or connected context. CortexDB eliminates this limitation by running four parallel retrieval channels.
How CortexDB thinks about hybrid retrieval
CortexDB treats retrieval as parallel reads against complementary indexes derived from the same immutable log. The four channels ensure that no context is missed:
- BM25 (Tantivy) handles exact terms, identifiers, and tokens.
- HNSW Vectors handles conceptual similarity.
- Graph Traversal handles connected context across entities.
- Cross-encoder Reranking evaluates the precise semantic relationship using Cohere rerank-v3.5.
Reciprocal Rank Fusion (RRF) combines the ranked lists into a single StratifiedPack. RRF requires no per-channel weight tuning and gracefully tolerates channels with differing score distributions.
What does each channel contribute?
Consider the query "what did Priya say about the SOC2 audit last week?" — a single sentence that requires every channel to land a correct answer.
| Channel | Hit on this query | Why it matters |
|---|---|---|
| BM25 (Tantivy) | Matches the literal terms Priya, SOC2, audit | Identifiers and acronyms (SOC2) survive lexically but get smeared in dense vectors |
| HNSW Vectors | Matches semantically-related messages (e.g. "the compliance review with Priya last Tuesday") | Captures paraphrases the lexical channel misses |
| Graph Traversal | Walks from entity Priya → edges typed mentioned/owns → recent compliance episodes | Surfaces connected context that mentions neither term but is causally linked |
| Cross-encoder rerank | Re-scores top candidates against the literal query | Resolves close-but-wrong matches; promotes the actual answer to position 1 |
The RRF fuser merges the ranked lists; the cross-encoder reranks the top-k of the fused list. The temporal phrase "last week" is resolved by the bi-temporal layer before retrieval runs, clipping every channel to the correct validity window.
What hybrid retrieval enables
- Comprehensive recall: Agents can query on identifiers and abstract concepts in the exact same prompt.
- Graceful degradation: BM25 and graph traversal still contribute meaningful context even if the vector channel returns nothing useful.
- Strict access boundaries: The retrieval engine seamlessly filters results across hierarchical scopes and respects the bi-temporal validity of every fact.
How CortexDB compares on retrieval architecture
Pinecone is a single-channel dense vector index. Mem0 wraps a vector index with LLM-rewritten memories but still relies on a single retrieval path. CortexDB achieves 93.8% on LongMemEval-S (beating Mem0 at 93.4%), and our approach prioritises architectural depth for complex enterprise environments because our architecture runs on an unmodified production write path.
| System | Lexical | Vector | Entity graph | Reranking | Fusion |
|---|---|---|---|---|---|
| CortexDB | BM25 (Tantivy) | HNSW vectors | Native traversal | Cross-encoder (Cohere) | Reciprocal Rank Fusion |
| Mem0 | None | Single dense vector | None | None | N/A |
| Zep | Partial | Dense vector | Coupled to write path | None | Custom |
| Pinecone | None | Dense vector | None | None | N/A |
| Neo4j | Text index | Optional vector | Native graph | None | Manual composition |
FAQ
What is 4-channel hybrid retrieval in CortexDB?
4-channel hybrid retrieval in CortexDB queries four complementary indexes during Cognitive Recall and fuses the ranked results. The channels are BM25 lexical search, HNSW vectors, graph traversal, and cross-encoder reranking.
What is Reciprocal Rank Fusion?
Reciprocal Rank Fusion (RRF) is a rank aggregation method that combines multiple ranked lists into one without requiring per-channel score normalisation. CortexDB uses RRF to fuse its retrieval channels cleanly.
Why is single-channel retrieval insufficient for agent memory?
Single-channel retrieval handles one query shape. Pure vector retrieval flattens identifiers, while pure lexical retrieval misses paraphrases. CortexDB uses multiple channels because agent queries never announce their shape in advance.
How does CortexDB compare to Pinecone for retrieval?
Pinecone relies on a single-channel dense vector index. CortexDB queries BM25, HNSW vectors, and an entity graph, reranks with a cross-encoder, and fuses everything seamlessly with RRF.
Does hybrid retrieval block the write path?
No. The retrieval indices are populated asynchronously from the event log. Heavy tasks like entity extraction run completely asynchronously and never block the main write path.