Memora: A Harmonic Memory Representation Balancing Abstraction and Specificity
4 hours ago
- #AI memory systems
- #long-horizon agents
- #information retrieval
- Memora is a scalable memory system for AI agents that decouples storage of rich content from lightweight retrieval mechanisms, balancing abstraction and specificity to enhance performance on long-horizon tasks.
- It introduces a dual-component memory entry with primary abstractions for indexing and cue anchors for flexible retrieval, enabling efficient information access without predefined ontologies.
- The system features a policy-guided retriever that actively refines queries and traverses cue anchors to uncover related memories, improving multi-hop reasoning and non-local context recall.
- Memora achieves state-of-the-art results on benchmarks like LoCoMo and LongMemEval, outperforming methods like RAG and Mem0 while reducing token usage by up to 98% and consolidating memory entries effectively.
- Future work includes MemLoop for self-improvement, Deferred Memory for optimized storage timing, and Group Memory for knowledge sharing, aiming to support long-term agent collaboration and organizational knowledge accumulation.