Long-term Memory for LLM-based Agents

Develop long-term memory mechanisms for large language model-based agents that enable reliable retention and retrieval over extended interactions and support continual learning within agentic systems.

Background

Within the discussion of LLMs as core components of advanced AI agents, the paper highlights current limitations in planning robustness, tool selection, and memory. In particular, the authors note that long-term memory capabilities are not yet adequately solved.

The text further emphasizes that existing paradigms do not readily support continual learning, underscoring the need for research into memory architectures and learning procedures that can operate over extended horizons and evolving tasks in agentic environments.

References

Additionally, long-term memory represents an open research problem, and the current paradigm does not readily support continual learning.

— Intelligent AI Delegation  (2602.11865 - Tomašev et al., 12 Feb 2026) in Section “Previous Work on Delegation” (LLMs paragraph)

These findings demonstrate the potential of continual learning at the harness level, while highlighting unresolved challenges in efficient retention evaluation, harness-content consolidation, and evaluation over longer interaction streams.

— Harness Continual Learning: Continual Adaptation Beyond Model Parameters  (2608.19013 - Kang et al., 19 Aug 2026) in Section Conclusion

This naturally raises a broader question: \textit{if semantic elaboration improves retrieval when applied at query time, can the same principle be incorporated during memory construction?}

— When Users Don't Ask: Benchmarking Context-Driven Memory Retrieval in Conversational Agents  (2609.03467 - Chang et al., 3 Sep 2026) in Section 6, subsection “Compression versus Elaboration in Memory Construction” (Section \ref{sec:compression-vs-elaboration})