Memory in the Age of AI Agents

TL;DR

A late-2025 survey of agent memory that reorganizes the field around a three-axis taxonomy — Forms (token-level / parametric / latent), Functions (factual / experiential / working), and Dynamics (formation / evolution / retrieval) — explicitly arguing the “long-term vs short-term” dichotomy that dominates earlier surveys is no longer the right partition. The accompanying paper list at https://github.com/Shichun-Liu/Agent-Memory-Paper-List is the practical artifact: a curated index of the recent agent-memory literature. The framing innovation is treating memory as orthogonal to context length and RAG — the survey distinguishes “agent memory” from the LLM-internal memory studied by the parametric-memory literature, from RAG (which is mechanism, not function), and from context engineering (which is provisioning, not memory).

Key claims

The “Forms” axis separates where memory lives: token-level (in-context text, RAG, summaries — externalized, human-readable), parametric (encoded in model weights via continual training, model editing, or distillation), and latent (carried in the forward-pass hidden state / KV cache, not persistent across sessions unless externalized) [§Forms taxonomy].
The “Functions” axis separates what role memory plays: factual (declarative knowledge — user profile, world state, ensuring cross-session consistency), experiential (procedural traces, success/failure patterns — what to do when), working (transient scratchpad for the current task) [§Functions taxonomy].
The “Dynamics” axis covers how memory changes: formation (how new items enter), evolution (how stored items consolidate, decay, or get rewritten), retrieval (how items are surfaced into context). Distinct from substrate decisions [§Dynamics taxonomy].
Explicitly distinguishes agent memory from three adjacent constructs: (a) LLM memory (parametric facts in a base model, no agentic loop), (b) RAG (one retrieval mechanism among many — a Forms-axis primitive, not a memory architecture), and (c) context engineering (prompt provisioning, not memory persistence) [framing §].
The companion paper list at https://github.com/Shichun-Liu/Agent-Memory-Paper-List is organized by the same taxonomy and tracks the recent surge in agent-memory work (MemGPT, Letta, mem0, Letta Filesystem, MAGMA, etc.) [GitHub README].

Method

This is a survey, not an empirical study. The authors curate the recent agent-memory literature, propose the three-axis Forms/Functions/Dynamics taxonomy, and populate each leaf with representative systems. The accompanying GitHub repo (https://github.com/Shichun-Liu/Agent-Memory-Paper-List) is the canonical reading list and is kept current via community PRs and issues.

Results

No quantitative results. The contribution is the reorganization itself: moving from temporal partitions (short-term vs long-term) and storage partitions (vector index vs text-record vs structural store, as in Rethinking Memory Mechanisms of Foundation Agents in the Second Half: A Survey) to a functional cross-product (Forms × Functions × Dynamics) that exposes design choices the older taxonomies collapsed together.

Why it’s interesting

This is the second comprehensive agent-memory survey filed in the last two weeks (the first being Rethinking Memory Mechanisms of Foundation Agents in the Second Half: A Survey from January 2026, with ~60 of the same author group). The two papers are in productive tension: 2602.06052 organizes by substrate × cognitive mechanism × subject (where the memory lives, what cognitive role, whose memory it is), while this paper organizes by Forms × Functions × Dynamics (where + what role + how it changes over time). The “Dynamics” axis is the new contribution — neither the older survey nor the parametric-memory literature treats memory evolution (consolidation, decay, rewriting) as a first-class axis. The taxonomy also draws a useful boundary against Parametric memory, which is the in-weights / in-state axis of memory studied independently of agentic loops — this survey explicitly claims agent memory is the functional layer that sits on top of that substrate.

For the team’s filed tool-use work, the Forms classification reads cleanly: Training-Free Group Relative Policy Optimization‘s natural-language experience library is token-level + experiential + evolution-heavy (Add/Delete/Modify/Keep on a text store); MiroThinker: Pushing the Performance Boundaries of Open-Source Research Agents via Model, Context, and Interactive Scaling‘s 256K-context recency retention is latent + working + retrieval-heavy; and Recursive Language Models‘s REPL-as-environment is token-level + working + formation-heavy (Python variables as fresh memory items).