# Reproduction: Memora: A Harmonic Memory Representation Balancing Abstraction and Specificity

## Pages

| Page |
| --- |
| [Executive summary](#/executive-summary) |
| [Claim 1: LoCoMo Policy Retriever reaches 0.863 LLM-as-a-Judge and exceeds full-context.](#/claim-1-locomo-policy-retriever-reaches-0-863-llm-as-a-judge-and-exceeds-full-context) |
| [Claim 2: LoCoMo Semantic Retriever reaches 0.849 LLM-as-a-Judge and exceeds full-context.](#/claim-2-locomo-semantic-retriever-reaches-0-849-llm-as-a-judge-and-exceeds-full-context) |
| [Claim 3: LongMemEval reaches 87.4% average accuracy with up to 98% token reduction.](#/claim-3-longmemeval-reaches-87-4-average-accuracy-with-up-to-98-token-reduction) |
| [Claim 4: Harmonic memory uses a primary abstraction, concrete memory value, and cue anchors.](#/claim-4-harmonic-memory-uses-a-primary-abstraction-concrete-memory-value-and-cue-anchors) |
| [Claim 5: RAG and KG retrieval are special cases of the Memora dual-layer design.](#/claim-5-rag-and-kg-retrieval-are-special-cases-of-the-memora-dual-layer-design) |
| [Conclusion](#/conclusion) |