agent-memory-survey

Author: Hao-Chien Lai DOI: 10.5281/zenodo.19551254 Date: April 2026

Abstract

The capability bottleneck of AI agents is shifting from reasoning to two intertwined challenges: what to remember and how to improve from experience. This survey provides a unified analysis of memory systems and self-iteration mechanisms in AI agents, covering 16 memory systems, 12 self-iteration frameworks, and 14 evaluation benchmarks spanning 2023–2026.

We organize memory systems into five families — context-resident compression, retrieval-augmented stores, reflective self-improvement, hierarchical virtual context, and policy-learned management — and map self-iteration mechanisms along three loop patterns: RL-driven, memory-driven, and reflection-driven.

Core thesis: Memory and iteration form a tightly coupled system. Memory quality determines the upper bound of iteration; iteration reshapes memory structure.

Key Findings

• Frozen Brain + Evolving Memory is among the most practical self-improvement architectures of 2026 — modifying the context is functionally equivalent to modifying the model's behavior, at a fraction of the retraining cost.

• Policy-learned memory management (F5) is a promising breakthrough direction — ALMA demonstrates that meta-learned memory designs outperform human-engineered baselines across four domains.