Microsoft's Memora lets AI agents remember with 98% fewer tokens
New memory system achieves 98% token reduction while beating Mem0 and RAG on benchmarks.
Deep Dive
Microsoft Research introduces Memora, an agentic memory framework for long-horizon AI agents, published at ICML 2026. It decouples what is stored (rich memory content) from how it is retrieved (lightweight abstractions and cue anchors) using short primary abstractions of 6–8 words paired with detailed memory values. Memora achieves state-of-the-art results on LoCoMo and LongMemEval, outperforming Mem0, RAG, and full-context inference while using up to 98% fewer context tokens. Open source on GitHub.
Key Points
- Memora uses a two-component structure: a 6-8 word primary abstraction for retrieval and a rich memory value for detail.
- Achieves up to 98% fewer context tokens than full-context inference while setting SOTA on LoCoMo and LongMemEval.
- Outperforms existing memory systems Mem0, RAG, and graph-based approaches (Zep, GraphRAG) on long-horizon agent tasks.
Why It Matters
Enables AI assistants to track multi-month projects with efficient, precise recall—no more stateless sessions or token waste.