Memory in the Age of AI Agents
by arXiv (Yuyang Hu, Shichun Liu, Yanwei Yue, Guibin Zhang, Boyang Liu et al.)
The reference taxonomy for agent memory — forms, functions, and dynamics — with benchmarks and frameworks.
Overview
Submitted to arXiv on 15 December 2025 (v2 revised 13 January 2026) as preprint 2512.13564, this survey argues that traditional long/short-term memory taxonomies no longer describe contemporary agent memory systems, and that loose terminology has obscured the field. It first delineates the scope of agent memory and distinguishes it from LLM memory, retrieval-augmented generation, and context engineering — a distinction most practitioners blur. It then analyzes memory through three unified lenses. By form, it identifies three dominant realizations: token-level, parametric, and latent memory. By function, it proposes a finer-grained taxonomy of factual, experiential, and working memory. By dynamics, it analyzes how memory is formed, evolved, and retrieved over time. For practical use it compiles a comprehensive summary of memory benchmarks and open-source memory frameworks, and it closes on emerging frontiers: memory automation, reinforcement learning integration, multimodal memory, multi-agent memory, and trustworthiness. The stated ambition is to serve as a conceptual foundation for treating memory as a first-class primitive in agent design rather than an add-on retrieval trick.
At a Glance
- Topic
- Agentic
- Level
- Advanced
- Format
- Paper
- Cost
- Free
- Duration
- ~2-3 hours for a full read
- Provider
- arXiv (Yuyang Hu, Shichun Liu, Yanwei Yue, Guibin Zhang, Boyang Liu et al.)
- Hands-on
- No
- Certificate
- None
What You’ll Learn
- ✓Where agent memory actually differs from RAG and from context engineering, and when each is the right tool
- ✓The forms taxonomy — token-level vs. parametric vs. latent memory — and the engineering trade-offs of each
- ✓The functional split between factual, experiential, and working memory in long-horizon agents
- ✓How memory is formed, evolved, and retrieved over time, including consolidation and forgetting dynamics
- ✓Which benchmarks to use when evaluating an agent memory system
- ✓Which open-source memory frameworks exist and what design choices each embodies
- ✓Open frontiers: memory automation, RL-integrated memory, multimodal and multi-agent memory, and trustworthiness risks
Highlights
- •Consolidates a fragmented 2025-2026 literature into one usable taxonomy
- •47 authors across the agent-memory research community
- •Includes a practical index of memory benchmarks and open-source frameworks, not just theory
- •Explicitly draws the agent-memory / RAG / context-engineering boundary that trips up production teams
Who It’s For
Best For
- ✓AI engineers designing long-horizon or persistent-memory agents
- ✓Teams choosing or evaluating an agent memory framework
- ✓Researchers entering agent memory who need a current map of the field
- ✓Anyone whose agent degrades over long multi-session interactions
Prerequisites
- •Solid grasp of LLM agents, tool use, and RAG
- •Comfort reading ML research papers
- •Some exposure to context-window and retrieval trade-offs in production
FAQ
What is Memory in the Age of AI Agents?
A 47-author survey that maps the fragmented field of agent memory into one coherent taxonomy, for AI engineers and researchers designing long-horizon agents. It separates agent memory from adjacent ideas like LLM memory, RAG, and context engineering, then catalogs the benchmarks and open-source frameworks that implement it.
Is Memory in the Age of AI Agents free?
Memory in the Age of AI Agents is free to access.
What level is Memory in the Age of AI Agents for?
Memory in the Age of AI Agents is aimed at a advanced audience. Recommended background: Solid grasp of LLM agents, tool use, and RAG, Comfort reading ML research papers, Some exposure to context-window and retrieval trade-offs in production.
How long does Memory in the Age of AI Agents take?
Expect roughly ~2-3 hours for a full read. Most learners work through it at their own pace.
What will I learn from Memory in the Age of AI Agents?
You'll learn: Where agent memory actually differs from RAG and from context engineering, and when each is the right tool; The forms taxonomy — token-level vs. parametric vs. latent memory — and the engineering trade-offs of each; The functional split between factual, experiential, and working memory in long-horizon agents; How memory is formed, evolved, and retrieved over time, including consolidation and forgetting dynamics; Which benchmarks to use when evaluating an agent memory system; Which open-source memory frameworks exist and what design choices each embodies; Open frontiers: memory automation, RL-integrated memory, multimodal and multi-agent memory, and trustworthiness risks.