Definition

Memora is a harmonic memory representation framework for long-horizon AI agents, published at ICML 2026 by microsoft Research. It decouples rich memory content from lightweight retrieval abstractions and cue anchors.

Key Points

  • Primary abstractions: Short phrases (6–8 words) index memory values; only abstractions are embedded for similarity search
  • Cue anchors: Context-aware tags providing alternative retrieval paths and cross-memory links
  • Retrieval policy: Markov Decision Process navigation beyond direct semantic similarity
  • Benchmarks: SOTA LoCoMo 86.3%, LongMemEval 87.4%; up to 98% fewer context tokens vs full-history baseline
  • Theory: Standard rag and knowledge-graph memory emerge as special cases
  • Open source: Code on GitHub; paper arXiv:2602.03315

Sources