Memora: A Harmonic Memory Representation Balancing Abstraction and Specificity
3 days ago
- #long-horizon tasks
- #memory systems
- #AI agents
- Memora is a scalable memory system designed to improve AI agent productivity on long-horizon tasks by decoupling storage from retrieval.
- It addresses the trade-off between specificity and abstraction in existing memory systems like Mem0 and RAG by using primary abstractions and cue anchors.
- Memora achieves state-of-the-art performance on benchmarks like LoCoMo and LongMemEval, reducing token usage by up to 98% compared to full-context inference.
- The system includes a policy-guided retriever that enables iterative, multi-hop reasoning, mimicking human-like memory recall.
- Future directions include MemLoop, Deferred Memory, and Group Memory to enhance learning, timing, and sharing of knowledge across agents.
- The research is published at ICML 2026, with code available on GitHub for community use.