Memora: A Harmonic Memory Representation Balancing Abstraction and Specificity

Agent memory systems must accommodate continuously growing information while supporting efficient, context-aware retrieval for downstream tasks. Abstraction is essential for scaling agent memory, yet it often comes at the cost of specificity, obscuring the fine-grained details required for effective reasoning. We introduce Memora, a harmonic memory representation that structurally balances abstraction and specificity. Memora organizes information via its primary abstractions that index concrete memory values and consolidate related updates into unified memory entries, while cue anchors expand retrieval access across diverse aspects of the memory and connect related memories. Building on this structure, we employ a retrieval policy that actively exploits these memory connections to retrieve relevant information beyond direct semantic similarity. Theoretically, we show that standard Retrieval-Augmented Generation (RAG) and Knowledge Graph (KG)-based memory systems emerge as special cases of our framework. Empirically, Memora establishes a new state-of-the-art on the LoCoMo and LongMemEval benchmarks, demonstrating better retrieval relevance and reasoning effectiveness as memory scales.

TToday's large language models reason impressively well in the moment, but they lack memory. Every conversation starts from scratch, leaving AI assistants unable to retain context across sessions or accumulate knowledge over time. Researchers have proposed many memory systems, from storing raw text fragments, to compressing experience into summaries, to organizing memory as a knowledge graph. But each design forces a tradeoff. Detail-preserving approaches fragment information into incoherent shards; abstraction-heavy approaches strip away the specifics the agent needs to act on. None of them give agents both at once. Our paper introduces Memora, a memory architecture that resolves this tradeoff by separating what is stored from how it is retrieved. Each memory entry has rich, expressive content (a project timeline, a multi-turn discussion) paired with two lightweight structural elements. A primary abstraction is a short canonical phrase that explains the main idea in the memory entry, capturing what the memory is fundamentally about. Cue anchors are a small set of contextual tags drawn from the memory's content, giving the agent alternative paths to the same entry from different angles. Only the primary abstraction and cue anchors are indexed and used for search. This decoupling lets related information consolidate into one coherent entry under its primary abstraction rather than fragmenting across many partial duplicates, and lets the agent reach the accurate memory through the abstractions from whichever angle a question suggests. Adding to this structure, Memora employs an active retrieval strategy. Rather than returning the most similar memories via similarity search, the retriever navigates the scaffolding. This lets the agent reach memories that are relevant but not directly similar to the question, capturing the kind of multi-step reasoning a person does when recalling connected events. On standard long-conversation benchmarks (LoCoMo and LongMemEval), Memora establishes new state-of-the-art results, demonstrating better retrieval relevance and reasoning effectiveness as memory scales.