Microsoft unveils Memora, a memory system for AI agents that cuts context use by up to 98%. Microsoft Research has presented Memora, a new memory architecture for artificial intelligence agents designed to let these systems remember information over long periods without having to continuously process full conversation histories.
According to the company, the system sets a new state of the art in long-term memory benchmarks such as LoCoMo and LongMemEval. Memora outperforms architectures such as RAG and Mem0, while cutting context token consumption by up to 98%.
Narrative coherence
The work will be presented at the ICML 2026conference, to be held from July 6 to 11 in Seoul (South Korea). The proposal addresses the ability to maintain persistent, contextual memory over weeks, months or even years of interaction in autonomous agents.
Memora's main innovation consists of separating the storage of information from the mechanism used to retrieve it. Instead of forcing the model to reread large amounts of text or rely on summaries that lose relevant information, the system stores the full content of the memory and organizes it through lightweight abstractions and contextual anchors that make later retrieval easier.
AI finds flaws in Windows and sparks a clash with Microsoft
According to Microsoft, this allows agents to maintain the narrative coherence of complex projects, remember past decisions, recall specific constraints and reconstruct the historical context of a task without having to indefinitely expand the context window of large language models.
The research reflects a shift in strategy in artificial intelligence development. Rather than racing to build models with ever-larger contexts, Microsoft is betting on persistent memory systems capable of accumulating experience and knowledge over the long term, a feature considered essential for the next generation of autonomous agents and enterprise AI assistants.