Infini Memory: Maintainable Topic Documents for Long-Term LLM Agent Memory

Researchers introduce Infini Memory, a novel persistent memory architecture for long-term LLM agents that organizes information as topic-structured documents rather than isolated records. The system consolidates observations through staged buffers and enables iterative evidence retrieval during inference, achieving 64.7% performance on MemoryAgentBench and demonstrating improved fact revision and memory maintenance capabilities.

Infini Memory addresses a fundamental challenge in AI development: enabling language model agents to maintain coherent, updatable memory across extended sessions. Current memory systems typically fragment information into disconnected records or static summaries, creating bottlenecks when agents need to reconcile contradictory facts or aggregate evidence across multiple interactions. This research reimagines memory architecture by treating it as semantic documents organized by topic, allowing related information to remain contextually connected.

The development reflects broader industry recognition that scaling AI agents requires rethinking data architecture from the ground up. As LLM applications move from single-session chatbots toward persistent autonomous agents operating over days or weeks, the limitations of traditional retrieval-augmented generation become apparent. Fragmented memory systems struggle with temporal coherence and fact evolution—critical capabilities for agents managing real-world information that changes over time.

Infini Memory's approach offers practical advantages for developers building agent systems. The topic-structured format maintains semantic coherence while the iterative retrieval mechanism (multiple tool calls rather than single-shot retrieval) allows agents to progressively refine understanding and resolve contradictions. The 64.7% benchmark performance validates the approach's viability, while ablation studies confirm that both structural organization and iterative inspection contribute meaningfully.

For the broader AI development community, this work signals that memory systems are becoming a competitive differentiator. Organizations developing production agents will need sophisticated memory architectures to handle real-world complexity. The research opens opportunities for specialized memory systems that could become infrastructure components for the next generation of autonomous agent platforms.

• →Topic-structured memory documents enable better fact revision and evidence aggregation compared to isolated record storage.
• →Iterative agent-driven retrieval through multiple tool calls outperforms single-retrieval approaches for complex memory access.
• →Infini Memory achieves 64.7% on MemoryAgentBench, demonstrating practical viability for long-term agent applications.
• →Staged consolidation of observations into coherent contexts reduces memory fragmentation and improves semantic coherence.
• →Memory architecture is emerging as a critical infrastructure challenge for scaling autonomous AI agents beyond single sessions.