AI agent memory is the set of mechanisms an agent uses to retain and retrieve information across interactions. If you are asking, “What is AI agent memory, and what types are there?”, the key distinction is that stored information is not automatically shown to the model: the system selects and assembles relevant material for each model call.
What AI agent memory does
A useful definition comes from the AWS Well-Architected Agentic AI Lens glossary: agent memory is “the mechanisms by which agents store and retrieve information across interactions.” That means memory is a system function, not necessarily one database. It includes decisions about what to keep, where to store it, how to find it later, and what to include in the model’s context.
As an Amazon Associate I earn from qualifying purchases.
Three terms describe different parts of that system. Short-term memory is recent, session-scoped state. Long-term memory is selected information that persists across sessions. Working memory is the context assembled for a particular model call. The first two describe retention and duration; working memory describes what the model receives now.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
Types of AI agent memory
Memory labels overlap because they classify different things. Short-term and long-term describe duration or scope. Semantic, episodic, and procedural describe the content being remembered. Working memory describes the assembled context, rather than a required durable store.
| Type | What it contains | Example | Design implication |
|---|---|---|---|
| Short-term (session) memory | Recent state for an active conversation or task, such as recent turns, tool results, and task variables | A tool result needed to finish the current request | Manage it against the session lifecycle and context limits; production services may need state outside a single process. |
| Long-term (persistent) memory | Selected information carried across sessions | A stable user preference or the outcome of an earlier decision | Choose what to extract, how to consolidate and retrieve it, and who owns it; retaining every transcript is not the same as maintaining useful memory. |
| Working memory | The context assembled for the current model call | Instructions plus relevant session details and retrieved persistent facts | Treat it as a context-assembly step, not necessarily a separate durable store. In Microsoft’s reference architecture, the model sees this assembled context, not the full underlying stores. |
| Semantic memory | Facts and attributes | A preference such as “prefers email” | Compact structured profiles or documents may suit stable facts; retrieve authoritative, changing domain information separately. |
| Episodic memory | Particular events or interactions, often with timestamps | A prior support interaction or decision | Retrieve relevant episodes on demand with suitable metadata and relevance filters rather than injecting an expanding history wholesale. |
| Procedural memory | Methods, workflows, or patterns learned through use | A method developed from repeated task outcomes | When an approved method already exists in a runbook, documentation, or code, use that authoritative source or tool instead of duplicating it as agent memory. |
The Microsoft multi-agent reference architecture’s Memory chapter, last updated August 4, 2026, puts the working-memory distinction plainly: “Working memory is the only thing the model ever sees. STM and LTM are design decisions about what gets to be there and at what cost.” In practice, the context may combine system instructions, the active conversation, selected tool outputs, and retrieved long-term items.
How an agent memory loop works
-
Capture the active state
Keep conversation turns, tool results, and variables needed for the current task in session memory. Google Cloud’s architecture guidance describes in-process state as a simple development approach and externalized state as a production pattern when requests may be handled by different instances. Examples it names include Memorystore for Redis and Firestore; it also notes a relational database option for the cited ADK service.
-
Select what should persist
Extract information likely to be useful beyond the current session, such as a durable preference, relevant fact, decision, or useful past episode. Do not make transcript retention the default substitute for deciding what matters.
The Tool Desk
Outbyte PC Repair FREEClear out junk files and repair common Windows errorsFree Scan →Outbyte Driver Updater FREEFix the driver behind crashes, sound loss and screen glitchesFind Drivers →Special offer. See more information about Outbyte and uninstall instructions. Please review EULA and Privacy policy. -
Consolidate and resolve changes
Merge duplicates, replace stale details where appropriate, and define how to handle conflicts—for example, whether a newer explicit correction overrides an older inferred preference. Microsoft Foundry documentation describes extraction, consolidation, and retrieval for its managed memory feature, which the page labels preview; availability and behavior may change.
-
Store information according to its purpose and scope
Choose representation and storage to match what must be recalled. Microsoft’s Memory Architecture Patterns, last updated August 4, 2026, describes structured relational or document profiles as a common fit for semantic facts and vector indexing for episodic recall. A graph is useful when traversing relationships justifies the added structure.
-
Retrieve only relevant items into working memory
Match retrieval to the current task, enforce access permissions, and fit selected material within the model’s context budget. Injecting every stored item can waste tokens and expose irrelevant information; retrieving too little can mean the agent misses useful context.
-
Apply lifecycle controls
Set boundaries for user, session, project, or tenant scope, and provide ways to correct, expire, or delete retained information. Shared enterprise content should be retrieved from permission-aware sources rather than silently copied into one user’s personal memory.
Recommended Free Tools
Special offer. See more information about Outbyte and uninstall instructions. Please review EULA and Privacy policy.
How to choose an architecture
There is no universally best memory design. Match storage and retrieval to the workload, the cost of a missed recall, and the sensitivity of the information.
| Decision | Option and when it fits | Trade-off to consider |
|---|---|---|
| Session state | In-process state keeps a development setup simple. Externalized state lets production instances retrieve and update session data across requests. | External state adds infrastructure and operational work; relying on one process can make state less suitable for multi-instance services. |
| Retrieval strategy | Push a compact profile into context when a small set of facts is broadly relevant. Pull specific memories on demand when relevance depends on the current task. | Push can consume tokens with details that do not matter for a request. Pull can add retrieval latency and miss information if the retrieval query or filters are poor. |
| Representation | Use structured records for stable facts, indexed event history for episodic recall, and procedural records for genuinely learned methods. | Do not force every kind of memory into a generic vector index; use authoritative runbooks, documentation, or code for already documented procedures. |
| Scope and access | Define whether a record belongs to a user, session, project, or shared context, and check permissions when retrieving shared material. | Loose boundaries can cause information from one project, channel, or tenant to surface in another. |
| Lifecycle | Set extraction thresholds, consolidation and conflict rules, retention periods, review options, and deletion behavior. | Without lifecycle rules, persistent records can become stale, duplicated, or difficult for users to correct. |
Evaluate operational quality with retrieval accuracy, precision and recall, token use, retrieval-plus-inference latency, and whether users have to repeat information. These measures expose different failures: fast retrieval is not useful if it returns the wrong memory, and high recall can still be costly if it injects too much context.
Agent memory versus a knowledge base or RAG
A practical distinction is ownership and authority. Agent memory captures information about a user, interaction, or collaboration that would otherwise be lost. A document repository, enterprise search index, or retrieval-augmented generation (RAG) corpus provides shared material that may change independently of any one conversation.
Retrieve authoritative shared content when needed and enforce permissions at retrieval time. Copying it into personal memory can create stale duplicates and blur who is allowed to see it. A vector database alone does not make a system’s contents “agent memory”; its role depends on what it stores and how the agent uses it.
Quick wins for a faster PC:
Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →Clear out junk files and repair common Windows errorsFree Scan →Why memory taxonomies differ
The categories above are practical design lenses, not a single settled taxonomy. The 2025 survey Memory in the Age of AI Agents describes other ways to classify memory: by form (token-level, parametric, or latent), function (factual, experiential, or working), and dynamics (how it is formed, changed, and retrieved). It also reports conceptual fragmentation and varying evaluation protocols across the field. Short-term and long-term labels alone therefore do not describe every modern system, and comparisons between systems require attention to how each defines and evaluates memory.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
