Title: DELTAMEM: Incremental Experience Memory for LLM Agents via Residual Trees

Abstract:

Large Language Model (LLM) agents are increasingly dependent on memory systems to derive insights from experiences accumulated during ongoing interactions. Nevertheless, treating these experiences as isolated, flat entities results in significant redundancy and retrieval conflicts. This occurs because similar episodes often contain overlapping information, while minor variations in scenes can lead to retrieved memories providing inconsistent or contradictory instructions. To mitigate these issues, we introduce the concept of "residual experience," which assumes that newly acquired knowledge typically represents an incremental modification of what is already known.

We present DeltaMem, a novel framework that structures experience memory into two distinct residual trees. The first tree catalogs goal-conditioned task experiences as reusable skills, while the second manages scene-level environmental knowledge. By utilizing a root node to store generalized base experiences and employing incremental delta nodes for subsequent variations, the system allows related experiences to share a common foundation, thereby eliminating duplication.

For the retrieval process, DeltaMem employs a failure-penalized similarity scan to identify the optimal match, reconstructing the complete experience through a composition of the root-to-match chain. Additionally, an autonomous consolidation mechanism distills high-frequency paths into new root nodes, facilitating a self-organizing structure that evolves from general heuristics to specialized variants. Our experiments across various interactive environments demonstrate that DeltaMem consistently surpasses existing baseline methods. To support further research, we have made the code available at https://github.com/import-myself/DeltaMem.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC