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

Why Agent Memory?

Large language models (LLMs) lose context when the interaction window is broken, which limits continuous dialogue and self‑evolution. An external memory that is readable, writable, expandable, and forgettable is required for agents to maintain long‑term state.

Formalization

The agent is modeled as a partially observable Markov decision process (POMDP). Memory is defined as a ternary operator that transforms raw interaction \(\phi_t\) into structured memory units \(M(\phi_t, \theta, \psi)\), where \(\theta\) denotes representation parameters and \(\psi\) denotes update rules.

Memory Forms

Token‑level : human‑readable text, JSON, or graph structures; low update cost; suitable for dialogue bots, legal audit, and any scenario requiring interpretability.

Parametric : fine‑tuned weights such as LoRA or adapter modules; not directly readable; medium update cost; used for role‑play, code generation, and other tasks where knowledge is embedded in model parameters.

Latent : KV‑cache, dense embeddings, or other vector stores; machine‑readable; minimal update cost; ideal for edge deployment, multimodal streams, and fast similarity search.

Memory Functions

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

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

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

Memory Dynamics

The full lifecycle follows a closed loop: Formation → Evolution → Retrieval .

Formation consists of five operations:

Semantic summarization of raw interactions.

Knowledge distillation into compact representations.

Structuring into schemas or graphs.

Latent encoding (e.g., embeddings, KV‑cache).

Parameter integration (e.g., LoRA adapters).

Evolution provides three primitives:

Consolidate – merge duplicates, de‑duplicate, and correct errors.

Update – modify existing memory units.

Forget – delete obsolete or privacy‑sensitive entries.

Retrieval follows four steps:

Trigger detection – decide when a query is needed.

Query construction – build textual or vector queries.

Retrieval strategy – nearest‑neighbor search, database lookup, or hybrid methods.

Post‑processing – ranking, compression, and plan generation.

Benchmarks and Open‑Source Frameworks

Thirty benchmark datasets covering memory, lifelong learning, and self‑evolution are collected. Representative open‑source systems include MemGPT, Mem0, Zep, and MemOS, compared on architecture, modality support, and scalability.

Relevant resources:

https://github.com/Shichun-Liu/Agent-Memory-Paper-List

https://arxiv.org/pdf/2512.13564

Future Directions (7 Frontiers)

Generative Memory – generate missing information instead of pure retrieval.

Automatic Memory Management – expose write/delete/update as callable tools for LLMs.

RL‑Driven Memory Strategies – replace hand‑crafted thresholds with end‑to‑end policy networks.

Multimodal Memory – unify video, audio, and sensor streams in a shared embedding space.

Multi‑Agent Shared Memory – role‑based, permissioned, privacy‑preserving shared stores.

World‑Model Memory – evolve from cache frames to queryable state simulators.

Trustworthy Memory – incorporate differential privacy, verifiable forgetting, audit logs, and GDPR‑compliant erasure.