Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Memora: A Harmonic Memory Representation Balancing Abstraction and Specificity

Menglin Xia, Xuchao Zhang, Shantanu Dixit, Paramaguru Harimurugan, Rujia Wang, Victor Ruhle, Robert Sim, Chetan Bansal, Saravan Rajmohan

TL;DR

Memora tackles the memory bottleneck in autonomous agents by balancing high-level abstractions with fine-grained details through a harmonic memory architecture. It introduces primary abstractions, cue anchors, episodic memory, and a policy-guided retrieval mechanism, with retrieval trained via Group-Relative Policy Optimization. Theoretical analysis shows Memora generalizes RAG and KG retrieval, while experiments on LoCoMo and LongMemEval establish state-of-the-art performance and substantial context efficiency. Overall, Memora provides a scalable, principled memory framework that enables robust, long-horizon reasoning with reduced token usage.

Abstract

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.

Memora: A Harmonic Memory Representation Balancing Abstraction and Specificity

TL;DR

Memora tackles the memory bottleneck in autonomous agents by balancing high-level abstractions with fine-grained details through a harmonic memory architecture. It introduces primary abstractions, cue anchors, episodic memory, and a policy-guided retrieval mechanism, with retrieval trained via Group-Relative Policy Optimization. Theoretical analysis shows Memora generalizes RAG and KG retrieval, while experiments on LoCoMo and LongMemEval establish state-of-the-art performance and substantial context efficiency. Overall, Memora provides a scalable, principled memory framework that enables robust, long-horizon reasoning with reduced token usage.

Abstract

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.
Paper Structure (58 sections, 5 theorems, 75 equations, 7 figures, 13 tables, 1 algorithm)

This paper contains 58 sections, 5 theorems, 75 equations, 7 figures, 13 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. Agentic Memory Management Systems
  4. Structured Memory Representations
  5. Method
  6. Problem Formulation
  7. Memora Overview
  8. Segmentation
  9. Episodic Memory
  10. Primary Abstraction
  11. Cue Anchors
  12. Policy-Guided Memory Retrieval
  13. Memory Retrieval Policy Formulation
  14. Group-Relative Policy Updates
  15. Theoretical Analysis
  16. ...and 43 more sections

Key Result

Theorem 4.1

Let $\mathcal{D}$ be a corpus and let $\mathcal{S}(\cdot)$ be a segmentation function that produces a set of chunks (segments). Consider a flat RAG retriever that, for any query $q$, returns where $\mathrm{sim}$ is a similarity function over chunk representations. Then there exists a Memora instantiation and a policy $\pi$ such that the retrieval set returned by Algorithm alg:policy_retrieval equ

Figures (7)

  • Figure 1: Overview of the Memora heterogeneous memory architecture.
  • Figure 2: Results for GRPO training.
  • Figure 3: Prompt for segmenting conversations into coherent episodic units.
  • Figure 4: Prompt for generating episodic memories from conversation segments.
  • Figure 5: Prompt for extracting factual memories from conversation segments.
  • ...and 2 more figures

Theorems & Definitions (11)

  • Theorem 4.1: Flat RAG as a Special Case of Memora
  • proof
  • Theorem 4.2: Implicit KG Retrieval as a Special Case of Memora
  • proof
  • Theorem 4.3: Explicit KG Retrieval as an Extended Case of Memora
  • proof
  • Definition 4.4: Retrieval classes
  • Theorem 4.5: Strictness under mixed-key constraints
  • proof
  • Theorem 4.6: Efficiency gain from abstraction-first + cue-anchor ANN retrieval
  • ...and 1 more