Memory in the Age of AI Agents
by arXiv (Hu, Liu, Yue, Zhang et al.)
The map of agent memory research: forms, functions and dynamics, plus every benchmark and framework.
Overview
"Memory in the Age of AI Agents" is a 47-author survey, led by Yuyang Hu, Shichun Liu, Yanwei Yue, Guibin Zhang and Boyang Liu, posted to arXiv on 15 December 2025 as arXiv:2512.13564, revised in January 2026 and filed under cs.CL and cs.AI. Its argument is that agent memory research has fragmented: papers published under the same banner differ substantially in motivation, implementation and evaluation protocol, and the traditional long-term versus short-term split no longer describes what contemporary systems do. The survey first draws a boundary — what counts as agent memory, and how it differs from LLM parametric memory, retrieval-augmented generation and context engineering — then reorganises the field along three axes. Forms covers the three realisations actually in use: token-level memory (text held in context or in an external store), parametric memory (knowledge written into weights), and latent memory (compressed hidden-state representations). Functions replaces long/short-term with a finer split into factual, experiential and working memory. Dynamics traces the lifecycle: how memories are formed, how they evolve, consolidate or are forgotten, and how they are retrieved. The paper then compiles memory benchmarks and open-source memory frameworks in one place — the section practitioners tend to use most — and closes on open frontiers: memory automation, reinforcement-learning integration, multimodal memory, multi-agent shared memory and trustworthiness. A companion GitHub repository, Agent-Memory-Paper-List, tracks 200+ papers under the same taxonomy, has over 2,000 stars and is actively updated.
At a Glance
- Topic
- Agentic
- Level
- Advanced
- Format
- Paper
- Cost
- Free
- Duration
- ~2-3 hour read; longer if you follow the companion 200+ paper list
- Provider
- arXiv (Hu, Liu, Yue, Zhang et al.)
- Hands-on
- No
- Certificate
- None
What You’ll Learn
- ✓Distinguish agent memory from LLM parametric memory, RAG and context engineering precisely
- ✓Classify a memory system by form: token-level, parametric or latent representation
- ✓Apply the factual, experiential and working memory taxonomy instead of long versus short term
- ✓Trace how memories are formed, consolidated, forgotten and retrieved over an agent's lifetime
- ✓Compare the current agent-memory benchmarks and what each one actually measures
- ✓Shortlist open-source memory frameworks against your own architecture's constraints
- ✓Identify open problems: memory automation, RL integration, multimodal and multi-agent memory
Highlights
- •Consolidates a field scattered across 200+ papers into a single navigable taxonomy
- •The companion GitHub paper list (2k+ stars) is kept current, so the survey does not go stale immediately
- •Explicitly separates agent memory from RAG and context engineering — the confusion most teams start with
- •Includes a practical inventory of benchmarks and open-source frameworks, not only theory
- •47 authors from across the agent-memory research community, with a January 2026 revision
Who It’s For
Best For
- ✓Researchers entering agent memory who need a map of prior work
- ✓Engineers designing a long-term memory layer for a production agent
- ✓Technical leads evaluating memory frameworks and benchmarks before committing
Prerequisites
- •Comfort reading machine-learning research papers and their notation
- •Working knowledge of LLM agents, retrieval-augmented generation and context windows
FAQ
What is Memory in the Age of AI Agents?
A large, current survey that maps the fragmented field of agent memory onto one taxonomy — forms (token-level, parametric, latent), functions (factual, experiential, working) and dynamics (formation, evolution, retrieval). It also compiles the available memory benchmarks and open-source frameworks, making it the fastest way to orient yourself before choosing or building a memory layer for an agent.
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: Comfort reading machine-learning research papers and their notation, Working knowledge of LLM agents, retrieval-augmented generation and context windows.
How long does Memory in the Age of AI Agents take?
Expect roughly ~2-3 hour read; longer if you follow the companion 200+ paper list. Most learners work through it at their own pace.
What will I learn from Memory in the Age of AI Agents?
You'll learn: Distinguish agent memory from LLM parametric memory, RAG and context engineering precisely; Classify a memory system by form: token-level, parametric or latent representation; Apply the factual, experiential and working memory taxonomy instead of long versus short term; Trace how memories are formed, consolidated, forgotten and retrieved over an agent's lifetime; Compare the current agent-memory benchmarks and what each one actually measures; Shortlist open-source memory frameworks against your own architecture's constraints; Identify open problems: memory automation, RL integration, multimodal and multi-agent memory.
Topics
Sources
This page was written from 3 sources, 2 on domains other than arxiv.org.