IterRet : Iterative Memory Reconstruction for Long-Horizon Conversational Agents (Small-Scale Frontier Track)

Modern agentic memory pipelines for long-horizon conversational agents are typically almost exclusively evaluated on multi-billion-to-frontier-scale backbones. We introduce IterRet, a novel iterative memory-reconstruction pipeline that unifies an associative memory graph, a closed-loop retrieval controller and a rubric-distilled experience bank to investigate whether such a genuinely new mechanism needs frontier scale to work at all. Every component of IterRet is a frozen-LLM prompt or a non-parametric structure, so it requires zero training FLOPs by construction. We evaluate it primarily on a frozen Qwen3-1.7B backbone reaching 17.72% token F1 on the LoCoMo long-horizon QA benchmark. The identical, unmodified pipeline on a frozen Qwen3-4B-Instruct backbone reaches 34.79% F1, outperforming recent state-of-the-art parametric-memory baselines. The consistent improvement across a ∼2.4× backbone-size change indicates the pipeline itself, rather than backbone scale, is responsible for the gains - evidence that a genuinely new active-reconstruction mechanism can be designed and shown effective at the small-scale frontier from the outset.