Selective Memory Retention for Long-Horizon LLM Agents

A new study introduces TraceRetain, a lightweight framework that gives frozen large language model agents a bounded external memory, finding that selective retention policies only outperform simple cache heuristics when the information stream contains noise [1]. The framework, detailed in a preprint submitted to arXiv on June 28, scores memory entries by interpretable features — including success, age, access frequency, redundancy, specificity, similarity, and downstream utility — and evicts the lowest-scoring entries when capacity is reached [1][2]. The work was authored by Pranath Reddy Kumbam [1]. Experiments were conducted in the ALFWorld environment using the gpt-5-mini model [1][2]. On clean benchmarks, external memory robustly improved task performance over a no-memory baseline across two random seeds, but the differences among various bounded retention policies fell within Wilson 95% confidence intervals [1][2]. The authors report that clean ALFWorld at T=100 to T=200 does not naturally exhibit the memory pollution that retention policies are designed to address [1][2]. To stress-test the system, researchers introduced a controlled noisy-write condition in which 75% of incoming entries were synthetic distractors [1][2]. Under this load, unbounded memory and a first-in-first-out policy limited to 50 entries, or FIFO-K50, saw Precision@5 scores degrade — from 20.2% to 12.4% and from 15.8% to 3.8%, respectively [1][2]. In contrast, the TraceRetain-CEM variant remained essentially unchanged, shifting from 16.9% to 16.6%, and preserved 97 out of 100 task successes [1][2]. The mechanism behind the degradation was traced to embedding-space similarity: unbounded memory exhibited the highest mean similarity score of 0.87 but the lowest precision, indicating that failed distractors were positioned close to the query in the embedding space [1][2]. In a held-out in-distribution evaluation, memory-augmented policies solved between 47 and 49 of 50 tasks, compared with 39 of 50 for the no-memory baseline [1][2]. The preprint appears on arXiv, an open-access repository that has hosted e-prints in physics, mathematics, computer science, and related fields since 1991 and now receives roughly 24,000 submissions per month [6]. The study falls within the computer science subfield of artificial intelligence and was posted under the cs.AI category [1].