Memory for Autonomous LLM Agents: Mechanisms, Evaluation, and Emerging Frontiers

Abstract

Memory systems enable LLM-based agents to persist and recall information across multiple interactions, transforming stateless text generators into genuinely adaptive systems. This survey formalizes agent memory as a write-manage-read loop tightly coupled with perception and action, establishing a structured approach to understanding memory architecture in autonomous systems. The work surveys memory design and implementation practices from 2022 through early 2026.

Key Contributions

• Core framework formalizing agent memory as a write-manage-read loop
• Three-dimensional taxonomy spanning temporal scope, representational substrate, and control policy
• Identification of 5 mechanism families implementing the write-manage-read loop
• Analysis of evaluation shift from static recall benchmarks to multi-session agentic tests
• Coverage of practical applications: personal assistants, coding agents, games, scientific reasoning
• Engineering considerations: filtering, contradiction handling, and privacy

Write-Manage-Read Loop

The fundamental abstraction organizing all agent memory systems:

1. Write — Encoding and storing new information from agent experiences
2. Manage — Organizing, consolidating, and maintaining stored memories over time
3. Read — Retrieving relevant memories to inform current reasoning and action

Five Mechanism Families

1. Context-Resident Compression

• Summarizing and compressing interaction history to fit within context windows
• Progressive distillation of episodic memories into compact representations
• Trade-off between information fidelity and context budget

2. Retrieval-Augmented Stores

• External memory stores (vector databases, knowledge graphs) accessed via retrieval
• Embedding-based similarity search for relevant memory retrieval
• Hybrid retrieval combining semantic similarity with temporal recency and importance

3. Reflective Self-Improvement

• Agents generating insights and lessons learned from past experiences
• Self-critique mechanisms identifying errors and updating behavioral policies
• Meta-cognitive processes distilling episodic memories into strategic knowledge

4. Hierarchical Virtual Context

• Multi-level memory hierarchies mimicking human short-term/long-term memory
• Working memory for active reasoning, episodic memory for experiences, semantic memory for facts
• Attention-based memory access prioritizing relevant information across hierarchy levels

5. Policy-Learned Management

• Learned policies for deciding what to store, when to consolidate, and how to retrieve
• Reinforcement learning for memory management optimization
• Adaptive memory strategies that evolve with agent experience

Open Challenges

• Continual consolidation — gracefully integrating new knowledge without catastrophic forgetting
• Causally grounded retrieval — retrieving memories based on causal relevance, not just similarity
• Trustworthy reflection — ensuring self-generated insights are reliable and grounded
• Multimodal embodied memory — extending memory systems to visual, spatial, and proprioceptive modalities