Rethinking Memory Mechanisms of Foundation Agents in the Second Half: A Survey

TL;DR

A 200+ paper survey that organizes the recent (2023 Q1–2025 Q4) explosion of foundation-agent memory work along three orthogonal axes: substrate (external vs internal), cognitive mechanism (episodic / semantic / sensory / working / procedural), and subject (agent-centric vs user-centric). The framing argument is that AI has entered its “second half” — long-horizon, multi-session, user-dependent deployments — and that memory, not modeling, is the binding constraint on real-world utility. The paper catalogues storage formats (vector index, text-record, structural store, hierarchical store, weight-internal, latent-state), agent-topology operations (single-agent vs multi-agent routing), learning policies that train memory management itself, scalability over context length and environment complexity, and benchmarks/metrics. It is a reference for what the agent-memory design space actually contains right now, rather than a new method.

Key claims

The field has shifted from prioritizing model architecture and benchmark scores to prioritizing real-world utility in long-horizon, dynamic, user-dependent environments, where context “explodes” and static, single-prompt mechanisms are insufficient [§1, §2.1].
218 key papers between 2023 Q1 and 2025 Q4 were curated from a larger pool; publication volume surged exponentially through 2025 with a peak in Q4 [§3, Fig. 3].
Agent memory can be taxonomized along three orthogonal axes: substrate (where information lives — external store vs network weights vs runtime activations), cognitive mechanism (functional role — episodic / semantic / sensory / working / procedural), and subject (who it serves — the agent vs the user) [§3, Figs. 2, 4].
External memory splits into four storage formats: vector index (HNSW/IVF/PQ over embeddings, RAG-style), text-record (human-readable core summary + episodic ledger + semantic list), structural store (relational tables, knowledge graphs, hierarchical trees with embedding+keyword traversal), and hierarchical store (dedicated per-function modules — core/episodic/semantic/procedural/resource — coordinated by a meta-memory manager) [§3.1.1].
Internal memory splits into weight memory (continual learning, model editing, distillation — durable but costly and prone to interference/overwriting) and latent-state memory (forward-pass hidden activations carried across steps — not persistent across sessions unless externalized) [§3.1.2].
External memory’s advantage is separation of computation from knowledge: updates without retraining, preservation of original form, modularity; its disadvantages are inference latency, retrieval reliability under poor ranking, and escalating storage/indexing costs as memory grows — requiring summarization, selective retention, forgetting, deduplication, and periodic pruning [§3.1.1].
The survey explicitly frames memory subject (agent-centric vs user-centric) as a design dimension prior surveys conflate — memory architectures optimized for the agent’s own task progress differ from those optimized for persistent per-user state across sessions [§1, §3.3 implied].
Memory operations under multi-agent topologies are distinct from single-agent operations and require routing primitives, not just read/write [§4 (referenced)].
Learning policies over memory operations (when to store, when to retrieve, when to forget) are an increasingly important axis — agents are now trained to manage memory rather than having memory management hard-coded [§5 (referenced)].
Six open challenges in foundation-agent memory are outlined to guide future design [§9 (referenced)].

Method

This is a survey, not a method paper. The authors collected a corpus by querying Google Scholar with memory-related keywords (agent memory, long-term memory, context management, personalization memory) and manually scanning major NLP/ML/IR/AI venues, then iteratively screened down to 218 articles between 2023 Q1 and 2025 Q4 [§3]. The taxonomy is presented as three orthogonal axes (substrate × cognitive mechanism × subject) plus four secondary dimensions: agent topology (single- vs multi-agent), learning policy (rule-based vs learned), scalability (over context length and environment complexity), and evaluation. Each leaf of the taxonomy is populated with representative implementations and citations.

Results

No quantitative results — this is a literature organization, not an empirical study. The headline empirical observations are:

Publication trend: exponential growth across 2023–2025, peaking in 2025 Q4 [§3, Fig. 3].
Corpus size: 218 curated papers from a several-hundred-paper initial pool [§3].
Substrate distribution (claimed but not quantified): vector-index RAG remains dominant for external memory; hierarchical-store designs (per-function modules + meta-manager) are gaining traction for long-horizon personalization; weight-based internal memory (continual learning, model editing, distillation) is an active but minority strand because of cost and interference [§3.1].

Why it’s interesting

The team has filed several agentic tool-use papers (MiroThinker, Training-Free GRPO, InfTool, RLM) but no dedicated memory survey — this paper is the closest thing to a map of the agent-memory design space and provides the vocabulary (substrate / mechanism / subject) for evaluating the memory architecture in each of those tool-use systems. In particular, Training-Free GRPO’s “natural-language experience library” is a text-record + agent-centric procedural memory under this taxonomy; MiroThinker’s 256K context + recency retention is latent-state + agent-centric episodic; RLM’s REPL-as-environment is closer to external structural store with a Python interpreter as the query engine. The survey itself is unlikely to be the source of a method, but its taxonomy is the cleanest available reference when arguing about which axis a new agent paper is actually moving on.