Memora: A Harmonic Memory Representation Balancing Abstraction and Specificity
TL;DR
Microsoft Research introduced Memora, a memory system for long-horizon AI agents that separates stored content from the way agents retrieve it. Instead of repeatedly loading full conversation history, Memora uses short primary abstractions and cue anchors as a lightweight access layer. Microsoft reports state-of-the-art results on LoCoMo and LongMemEval: 86.3% LLM-judge accuracy on LoCoMo, 87.4% on LongMemEval, and up to 98% fewer context tokens than full-context inference.
Nauti's Take
The interesting move here is not a bigger context window, but a better memory architecture. Memora’s core idea is simple: store rich detail, retrieve through compact and intentional access points.
That is closer to real work than the usual RAG drawer full of text fragments. Still, the proof is not the blog chart; it is messy production use with contradictory updates, privacy boundaries, and stale information.
Briefingshow
Today’s agents often fail less because of reasoning and more because of memory: they lose decisions, constraints, and earlier rationale. Memora targets that gap by keeping detailed memories while making retrieval cheaper and more structured. If it holds up beyond benchmarks, it could become a building block for project, support, and research agents that operate over weeks instead of minutes.
Sources
