Which Memory Operation Drives Recovery? A Factorial Study of Retrieve, Write, and Manage Adaptation under Domain Shift

Zhaoxiang Feng, Mingyang Yao, Charlie Sun, David Scott Lewis

Keywords: lifelong agents, memory adaptation, domain shift, Thompson sampling, factorial ablation, retrieval-augmented generation

TL;DR: A factorial study isolating retrieve, write, and manage adaptation in LLM agent memory under domain shift finds that store management yields the largest in-distribution gains, while post-shift recovery is dominated by base model competence.

Abstract: Memory-augmented LLM agents use external stores to accumulate and reuse experience across lifelong deployment, yet existing work treats memory as a monolithic system, adapting retrieval while leaving write and manage policies fixed. We argue that memory adaptation should be studied as a factorial problem across three coupled operations: retrieving stored experience (PROVIDE), writing new experience (TAKE-IN), and maintaining store quality (MANAGE). We introduce OMAC, an online memory architecture controller that uses block-level Thompson sampling over the joint configuration space of all three operations, and a factorial ablation framework that isolates each operation's contribution by selectively enabling or disabling its adaptation. Through controlled factorial experiments on a streaming domain shift, we find that adaptive store management consistently outperforms retrieval-only adaptation for in-distribution learning, challenging the prevailing assumption that retrieval is the primary lever for memory-augmented agents. We further show that the value of memory adaptation is modulated by domain difficulty: when the base model is already competent, no adaptation strategy provides meaningful lift, whereas management adaptation yields substantial gains when the model struggles. These findings establish that the memory lifecycle deserves the same systematic, per-operation analysis that has been applied to retrieval, and that store management is an underexplored yet high-impact axis of improvement for deployed agents.

Submission Number: 226