Infini Memory proposes topic-structured documents for LLM agent memory
Infini Memory is a persistent memory architecture for long-term LLM agents that organizes memory as topic-structured documents, achieving a 64.7% overall score on MemoryAgentBench.
Score breakdown
The paper addresses a core limitation of existing LLM agent memory systems — difficulty with evidence aggregation and fact revision across sessions — by introducing a structured, maintainable architecture that improves both how memory is organized and how it is retrieved.
- 01Infini Memory is proposed by Suozhao Ji, Baodong Wu, and Zehao Wang.
- 02The architecture organizes agent memory as topic-structured documents rather than isolated records, summaries, or indexed fragments.
- 03Each topic document collects related evidence, preserves metadata, and supports fact revision over time.
Suozhao Ji, Baodong Wu, and Zehao Wang introduce Infini Memory, a maintainable text-based persistent memory architecture designed to address shortcomings in how long-term LLM agents store and retrieve information across sessions. Existing approaches — storing observations as isolated records, summaries, or indexed fragments — struggle with evidence aggregation, fact revision, and ongoing memory maintenance. Infini Memory instead treats agent memory as a collection of topic-structured documents, where each topic document serves as a semantic unit responsible for gathering related evidence, preserving metadata, and updating facts as new information arrives.
The architecture introduces a two-stage write process: new observations are first staged in a buffer and then periodically consolidated into coherent textual contexts within the appropriate topic document.
The architecture introduces a two-stage write process: new observations are first staged in a buffer and then periodically consolidated into coherent textual contexts within the appropriate topic document. At inference time, rather than relying on a single retrieval step, an agentic retrieval procedure allows the LLM to read memory through iterative tool calls, enabling more targeted and flexible evidence inspection. Evaluated on MemoryAgentBench, Infini Memory achieves a 64.7% overall score. Ablation studies demonstrate that topic-structured maintenance and iterative evidence inspection each improve distinct, complementary aspects of long-term memory performance.
Key facts
- 01Infini Memory is proposed by Suozhao Ji, Baodong Wu, and Zehao Wang.
- 02The architecture organizes agent memory as topic-structured documents rather than isolated records, summaries, or indexed fragments.
- 03Each topic document collects related evidence, preserves metadata, and supports fact revision over time.
- 04New observations are staged in a buffer and periodically consolidated into coherent textual contexts.
- 05At inference time, an agentic retrieval procedure uses iterative tool calls instead of a single retrieval step.
- 06Infini Memory achieves 64.7% overall score on MemoryAgentBench.
- 07Ablations show topic-structured maintenance and iterative evidence inspection improve complementary aspects of long-term memory.
Topics
Summary and scoring are generated automatically from the original article. We always link back to the publisher and never republish images or paywalled content. Last processed Jun 10, 2026 · 15:34 UTC. How this works →