Memora: A Harmonic Memory Representation Balancing Abstraction and Specificity
Summary
Microsoft Research's Memora introduces a harmonic memory representation that decouples stored content from retrieval, enabling long-horizon AI agents to maintain context with far fewer tokens. The approach uses a two-part memory entry (primary abstraction and memory value) plus cue anchors and a policy-guided retriever to support multi-hop recall without a fixed ontology. On LoCoMo and LongMemEval benchmarks Memora achieves state-of-the-art results with up to 98% fewer tokens, signaling potential for scalable enterprise AI workflows.