Long Term AI Memory by creator of Apache Cassandra

cortexdb.ai

CortexDB is the long-term memory layer for AI systems — The problem is fundamental: today's AI agents are stateless. Every conversation starts from zero. The dominant approach to giving AI memory — having an LLM rewrite and merge your data on every single write — is lossy, fragile, and ruinously expensive. The LLM decides what to keep and what to throw away, replaces the original with a summary, and that decision is irreversible. Information it deemed unimportant today may be exactly what a future query needs tomorrow. CortexDB takes a fundamentally different approach: every piece of information is appended to an immutable event log and never overwritten. A lightweight LLM extracts entities and relationships asynchronously, but the original data is always preserved — if the extraction misses something, the raw event is still there for any future query or reprocessing. From this event stream. CortexDB automatically builds a temporal knowledge graph — entities, relationships, causal chains, and provenance — and uses hybrid retrieval combining vector search, full-text matching, graph traversal, and adaptive ranking to assemble the exact context an AI agent needs at query time. The results are not incremental. In controlled benchmarks using identical language models, identical embeddings, and identical test data across five production-scale scenarios, CortexDB achieved a huge gap that is structural, not incidental, because you cannot retrieve information you've already destroyed. The cost difference is equally dramatic because CortexDB's write path uses a lightweight extraction model while rewriting systems burn expensive LLM inference to merge and regenerate entire memory stores on every write operation.

CortexDB scales the same way Cassandra scales — through consistent hashing, partition-aware data placement, and leaderless replication, where every index, every graph shard, and every vector store is scoped to a partition from day one. Adding capacity means adding a node; the cluster rebalances automatically with zero downtime. A single-node deployment is simply a distributed system with one node — the same code path runs whether you have one machine or a hundred. This is not a single-node prototype that will be distributed later. Distribution is the architecture itself, at scale — retrofitting distribution onto a monolithic design costs more than building it right from the start. CortexDB is not a better version of what exists. It is a new layer of infrastructure — the memory layer — built from first principle that scales infinitely unlike any other solution in the market.