NanoMem: Iterative Evidence Synthesis for Temporal-Causal Reasoning in Agent Memory

Abstract

from personal histories spanning multiple sessions. A central challenge is the evidence addressability gap: the retrieval cue needed for the final answer may not be present in the original query, and must instead be inferred from intermediate evidence discovered in earlier retrieval rounds. Existing approaches either conflate session time with the time at which the described event occurred, or lack a mechanism to carry temporal inferences across retrieval steps to reach indirectly addressable evidence. This thesis investigates how a long-term conversational memory system can construct compact and temporally grounded evidence at query time, without relying on heavy write-time processing that fixes interpretations before the future query is known. I introduce NanoMem, a memory evidence synthesis framework that reformulates temporal memory search as iterative evidence construction. The core mechanism is a Temporal Evidence Pool that accumulates structured event records with separate event-time and session-time fields. At ingestion, NanoMem applies lightweight deterministic normalization to resolve relative temporal expressions. At search time, a Planner decomposes the query into retrieval cues, and a trainable Synthesizer extracts temporally grounded events from retrieved sessions and judges whether the accumulated evidence is sufficient to answer the query. When it is not, the inferred events feed back to the Planner to drive the next retrieval round. The Synthesizer is trained with GRPO using multi-level rewards covering answer correctness, evidence compactness, output format, and a per-round sufficiency verdict that provides direct process supervision for the stop-or-continue decision. A new diagnostic benchmark, TimeMemEval, is introduced to evaluate hidden-dependency temporal bridge search, where the answer-bearing evidence is reachable only after an intermediate temporal anchor has been established. NanoMem is evaluated on LoCoMo, LongMemEval, and TimeMemEval. On all three benchmarks, NanoMem achieves the best accuracy among all evaluated methods, improving over the strongest prior baselines by 12.5, 11.8, and 31.0 percentage points, respectively. GRPO training also reduces returned evidence length by 59.5% and 66.3% on LoCoMo and LongMemEval relative to the untrained variant, while simultaneously improving accuracy. The results support the thesis position that query-time evidence synthesis, structured temporal grounding, and reward-aligned sufficiency judgment together address the evidence addressability gap more effectively than full-context prompting or write-time memory construction.