Memora: Microsoft Research's Harmonic Memory for Long-Horizon Agents
Source: Microsoft Research — 2026-06-29
Summary
Microsoft Research published Memora, a long-term memory framework for AI agents that separates each memory into a short "primary abstraction" used for similarity search and a rich, uncompressed "memory value" for retrieval — avoiding the usual trade-off between compact indexing and preserving fine-grained detail. A policy-guided retriever treats recall as an iterative reasoning process, refining queries and hopping through "cue anchor" tags rather than doing a single top-k semantic lookup. The paper, headed for ICML 2026, reports state-of-the-art results on the LoCoMo and LongMemEval long-context benchmarks while storing far fewer entries and tokens than comparable systems like Mem0.
Key Takeaways
- Each memory stores two representations: a 6–8 word "primary abstraction" for search, and a detailed "memory value" that stays uncompressed — new information about an existing topic consolidates into the same abstraction instead of spawning duplicate entries.
- "Cue anchors" are context-derived tags that give the retriever alternate paths to related-but-not-semantically-similar memories, avoiding a fixed ontology.
- Retrieval is iterative and policy-guided rather than one-shot top-k: the system refines its query and expands through cue anchors, mimicking multi-hop human recall.
- On LoCoMo (600-turn dialogues), Memora scores 86.3% LLM-judge accuracy, beating RAG, Mem0, Nemori, Zep, LangMem, and full-context inference; on LongMemEval (115K-token contexts) it scores 87.4%, with its biggest gains on multi-hop reasoning.
- Efficiency: roughly 50% fewer stored entries than Mem0 (344 vs. 651) and up to 98% fewer tokens consumed than reloading full conversation history.
- Code is released at github.com/microsoft/Memora alongside the ICML 2026 paper.