Seamless: Self-Maintaining Knowledge Graph Memory for Retrieval-Augmented Language Models

Abstract

Large language model (LLM) agents have bounded context windows and therefore need external storage to retain knowledge across long interactions. Keeping that knowledge consistent requires reconciling changing facts, duplicate mentions, and competing retrieval priorities. We present Seamless, a self-maintaining knowledge-graph (KG) memory that coordinates note-level triage, decay-aware retrieval, and an autonomous editor through read-only recall, post-session updating, and periodic reformulation. Its contribution is the coordination of these mechanisms, rather than the invention of graph memory or memory editing individually. Under a shared embedding model and response generator, Seamless scores 6.563/10 on LongMemEval Medium, exceeding Mem0 by 0.713 points (95% confidence interval [0.620, 0.801]). Disabling the editor reduces its score by 0.598 points [0.530, 0.665]. A blinded human audit of 300 questions and 1,200 responses supports the LongMemEval ordering. On LoCoMo, where Seamless and Mem0 both score near ceiling, their difference is small (1.833 percentage points [0.167, 3.500]) and is not resolved on the audited subset. The difference between simple and pulsed decay is unresolved. Measured mean query latency is 1.539 seconds, with a further 0.541 seconds of maintenance amortized per query. The findings support coordinated maintenance while identifying limits in verbatim recall, temporal serialization, and consolidation cost.