What Is Agentic Memory?
Agentic memory is the set of mechanisms that let an AI agent keep information from its own work and use it in later steps, sessions or tasks, because the language model underneath retains nothing between calls. The term is used in two ways. Broadly, it means any memory an agent uses, usually divided into working, episodic, semantic and procedural memory. Narrowly, it means memory the agent manages itself, deciding what to store, link, update or discard.
Updated
What is agentic memory?
The vocabulary comes from cognitive science. The CoALA framework (Sumers, Yao, Narasimhan and Griffiths, TMLR 2024) starts from the premise that "Language models are stateless: they do not persist information across calls." It organizes a language agent's memory into short-term working memory and three long-term stores: episodic, semantic and procedural. The split mirrors cognitive architectures such as Soar, which CoALA describes as building on psychological theories.
The narrow sense is newer. A-MEM: Agentic Memory for LLM Agents (Xu et al., NeurIPS 2025) describes a memory that organizes itself "in an agentic way." Following the Zettelkasten method, each new memory becomes a linked note with "contextual descriptions, keywords, and tags" and can "trigger updates to the contextual representations and attributes of existing historical memories." AgeMem (Yu et al., 2026) "exposes memory operations as tool-based actions," so the agent decides "what and when to store, retrieve, update, summarize, or discard information." Anthropic's context engineering guide uses the term more loosely, for structured note-taking.
Agent Memory: Architecture, Types, and Trade-Offs covers the memory types in depth, and AI Agent Memory Architecture makes the case for keeping live operational state separate from memory.
How agentic memory works
| Type | What it holds | Typical home | How it changes |
|---|---|---|---|
| Working | The current step's inputs, goals and results | Context window or runtime state | Every step |
| Episodic | What happened, in order | Append-only log, often embedded | Appended, rarely revised |
| Semantic | Facts, preferences, domain knowledge | Keyed records plus a vector index | Revised as facts change |
| Procedural | How to act | Instructions, code, model weights | Edited deliberately |
A memory system runs a loop over these stores. It writes, often by having a model extract facts from a finished interaction. It organizes, by linking related records and superseding outdated ones. It retrieves: in Generative Agents (Park et al., 2023), memories are scored on recency, importance and relevance. And it forgets what no longer earns its place. Writes are not equally safe: CoALA calls writing to procedural memory "significantly riskier than writing to episodic or semantic memory."
Conflicts are the everyday problem. A Salesforce Engineering post on the memory system behind its Agentforce agents notes that "Enterprise systems may contradict conversational signals, and memory must represent uncertainty rather than false certainty" (Salesforce Engineering).
Why agentic memory matters in real-time systems
Most memory designs assume one agent serving one user; the authors of Collaborative Memory note that most approaches "assume a monolithic, single-user context." Private memory tolerates some staleness, because only one agent writes to it. That changes when several agents read and write the same memory, such as a customer record that a voice agent and a chat agent both update, or a running total of what purchasing agents have committed against one budget. The memory is now shared mutable state, with the failure mode of any shared state: two agents read the same remaining amount, both act, and the budget is exceeded.
That paper's framework centers on permissions, with private and shared tiers and provenance on each fragment. Access control decides who may read a memory. It does not decide what happens when two agents write to it at once; that takes concurrency control in the store. A second rule applies even to one agent: a recalled balance or limit is a record of an earlier read, not the current value. Read anything a decision gates on from the system that owns it, when the decision runs. One agent remembering one user's preferences needs none of this; a vector index, or Postgres with pgvector, handles that retrieval.
FAQ
Related terms
Episodic memory in AI agents is a time-ordered record of the agent's own past events, actions and outcomes, retrieved to reuse experience in later decisions.
Semantic memory is long-term memory for facts and concepts. In AI agents it is the store of facts an agent keeps about the world, its users and itself.
Procedural memory is an AI agent's knowledge of how to act: the model weights, code, instructions and skills that set its behavior. The term is from psychology.
Context offloading moves information out of an LLM's context window into external storage, such as files or notes, and pulls it back only when a step needs it.
Multi-writer contention is two or more actors reading and writing the same mutable state in the same short window. It is not multi-tenancy or partitioning.
Shared context means every service and agent in a decision system reads one internally coherent snapshot of state, instead of private copies that diverge.