Memora scales agent memory to boost long-horizon productivity

Abstraction: Harmonic memory system decoupling storage and retrieval for long-horizon agents

Key points:

• Memora decouples what is stored (rich memory values) from how it is retrieved (lightweight primary abstractions + cue anchors), resolving the abstraction-specificity tension in prior systems like Mem0 and RAG
• Primary abstraction is a 6-8 word phrase capturing the memory's essence; cue anchors are short context-aware tags providing alternative retrieval paths to the same memory entry
• Policy-guided retriever iteratively refines queries and expands through cue anchors for multi-hop reasoning rather than single-shot top-k retrieval
• Achieves state-of-the-art on LoCoMo (86.3%) and LongMemEval (87.4%), outperforming Mem0, RAG, Zep, LangMem, and full-context inference
• Uses 98% fewer tokens than full-context inference and stores roughly half the entries per conversation compared to Mem0 (344 vs. 651)
• Published at ICML 2026; code released at github.com/microsoft/Memora

Connections: Microsoft · AI Agents · Agent Memory · Retrieval Augmented Generation

Source: https://www.microsoft.com/en-us/research/blog/memora-a-harmonic-memory-representation-balancing-abstraction-and-specificity/