Researchers map memory systems for AI agents — framework distinguishes three types (token, parametric, latent) to clarify fragmented field

A team of 45+ researchers published a comprehensive survey on agent memory (how AI systems recall and use information over time) on January 13, 2026, identifying three dominant memory architectures: token-level (raw text stored in input), parametric (learned into model weights), and latent (compressed vector representations). The paper clarifies that agent memory differs from general LLM memory and RAG (retrieval-augmented generation — looking up facts in external databases).

The survey proposes a three-layer taxonomy: factual memory (storing facts), experiential memory (learning from past agent actions), and working memory (holding current task context). This replaces vague concepts like 'short-term' and 'long-term' with precise definitions of what each memory type does, letting engineers understand which approach fits their use case.

For AI developers and companies building autonomous agents (AIs that make decisions and take actions without human intervention at each step), this framework ends confusion about which memory design works best for different tasks — whether you need to recall a conversation from last week, learn from mistakes, or track what you're currently working on. Teams can now reference standardized benchmarks and open-source memory frameworks instead of reinventing solutions.