Memora: A Harmonic Memory Representation Balancing Abstraction and Specificity
| Source: Microsoft Research Blog
Tags: Memora, Microsoft Research, ICML 2026, agent memory, RAG, long-horizon agents, context efficiency
Microsoft's Memora separates memory storage from retrieval in AI agents, achieving state-of-the-art on LoCoMo and LongMemEval benchmarks while using 98% fewer context tokens than full-context inference — directly addressing the memory bottleneck limiting long-horizon agent deployments.
Details
Today's AI agents are effectively stateless: every session starts from zero, and handling long conversations forces re-reading the entire history. Microsoft Research's Memora, published at ICML 2026 with code on GitHub, addresses this by decoupling what is stored from how it is retrieved. Rich memory content — project timelines, multi-turn discussions — is indexed via lightweight abstractions called cue anchors, so retrieval stays efficient as memory grows. On established benchmarks, Memora sets new SOTA on LoCoMo and LongMemEval, outperforming Mem0, RAG, and graph-based systems like Zep and GraphRAG. The key metric: up to 98% fewer context tokens compared to full-context inference while maintaining comparable or better recall quality. The framework scales as task complexity grows, making it practically relevant for enterprise agents running multi-month workflows. For practitioners building long-horizon agents, Memora offers a concrete architecture worth evaluating. The open-source release on GitHub means it can be integrated into existing agentic frameworks. The main open question is how it performs across domains beyond the specific benchmark settings.