Unlocking Agent Memory: A Comprehensive Survey of Forms, Functions, and Dynamics

Why Agent Memory?

Large language models (LLMs) lose context when a conversation is interrupted, leading to a "goldfish brain" effect. To enable agents that can interact continuously, evolve autonomously, and retain knowledge, a persistent, readable, writable, and forgettable memory component is essential.

Formalizing Agent Memory

Agents are abstracted as partially observable Markov games. Memory is defined as a triple operator consisting of:

Formation (F) : Transform raw interactions \(\phi_t\) into memory units.

Evolution (E) : Merge, deduplicate, correct, and forget stored information.

Retrieval (R) : Query the memory on demand.

Memory Forms (What Does Memory Look Like?)

The survey identifies three primary memory forms and further refines token‑level memory into three structural variants.

Token‑level : Text, JSON, or graph representations; human‑readable; low update cost; typical for dialogue bots and legal audit.

Parametric : LoRA or adapter weights; not human‑readable; medium update cost; used in role‑play and code generation.

Latent : KV‑cache or embeddings; machine‑readable; negligible update cost; suited for edge deployment and multimodal streams.

Memory Functions (What Is Memory Used For?)

Agent memory serves three major functions:

Factual Memory – Stores "what I know": user profiles, document states, world knowledge.

Experiential Memory – Stores "what I learned": success/failure trajectories, strategies, executable skills.

Working Memory – Stores "what I think now": single‑turn compression, multi‑turn state folding, planning cache.

Memory Dynamics (How Does Memory Operate?)

The full lifecycle forms a closed loop: Formation → Evolution → Retrieval . Key operations include:

Formation : Semantic summarization, knowledge distillation, structuring, latent‑space encoding, parameter internalization.

Evolution : Consolidate, update, and forget mechanisms.

Retrieval : Trigger timing, query construction, retrieval strategy, post‑processing.

Resources (Benchmarks & Open‑Source Frameworks)

The survey compiles more than 30 benchmark datasets for memory, lifelong learning, and self‑evolution, and compares over 20 open‑source frameworks such as MemGPT, Mem0, Zep, and MemOS.

Frontier Topics (Seven Emerging Directions)

Generative memory surpassing retrieval‑based approaches.

Automatic memory management as callable tools for agents.

Reinforcement‑learning‑driven memory strategy optimization.

Multimodal memory integrating video, audio, and sensor streams.

Shared memory across multiple agents with role‑based access control.

World‑model memory transitioning from cache frames to queryable simulators.

Trustworthy memory incorporating differential privacy, verifiable forgetting, audit logs, and GDPR‑compliant erasure.

Memory in the Age of AI Agents: A Survey Forms, Functions and Dynamics
https://github.com/Shichun-Liu/Agent-Memory-Paper-List
https://arxiv.org/pdf/2512.13564