Iterative Memory Reconstruction for Long-Horizon Conversational Agents

Answering queries over multi-session conversations requires retrieving and chaining evidence distributed across a history far larger than any single context window. Standard retrieval-augmented approaches retrieve memory passively in a single pass, failing whenever a correct answer requires assembling evidence across non-adjacent turns. We present IMR (Iterative Memory Reconstruction), a framework that unifies three complementary mechanisms into a single active reconstruction loop: (i) a layered associative memory graph that exposes semantic facets as intermediate bridges between query anchors and memory content, supporting semantically guided multi-step traversal with explicit path pruning; (ii) a closed-loop controller with persistent evidence-gap tracking that interleaves retrieval and reflection steps and terminates only when gaps are resolved or a budget is exhausted; and (iii) an offline experience bank that scores historical search trajectories step by step and distills reusable planning and reflection guidance, steering subsequent inference away from known failure modes. Using a frozen 4B-parameter backbone with no fine-tuning, IMR achieves a token F1 score of 34.79% on LoCoMo at 4.96 retrieval iterations per question, outperforming well-known baselines on the most evidence-intensive question types. These results demonstrate that active, iterative memory reconstruction is both tractable and semantically effective.