AI agents use several distinct types of memory — in-context, external, episodic and semantic — to maintain state and recall relevant information across steps and sessions. Understanding these types shapes what your agents can reliably do.
AI agents do not have persistent memory in the way a human does. By default, each inference call starts with only what is in the current context window — when the interaction ends, nothing is retained. For agents to operate effectively across multi-step tasks, sessions and time, memory must be explicitly designed and managed: information written to external stores, retrieved when needed, and expired when stale. Understanding the distinct types of agent memory — and what each is suited for — is foundational to designing agents that are both capable and reliably accurate.
Agent memory is typically categorised into four types, each serving a different function:
In-context memory is everything currently in the model's active context window — the conversation history, retrieved documents, tool outputs and the ongoing reasoning trace. It is temporary, erased at the end of the session, and bounded by the model's context window limit.
External (persistent) memory is information written to and retrieved from a database, file system or object store between sessions. An agent might record a completed task summary, a user's stated preferences, or the current state of a long-running workflow. On subsequent invocations, the agent retrieves this stored state to resume from where it left off.
Episodic memory is a record of specific past interactions — structured logs of what the agent encountered, what it did, and what the outcome was. This is retrieved to inform current decisions based on relevant past experience rather than general knowledge.
Semantic memory is general knowledge or facts stored in a vector database and retrieved by similarity search. This is the memory type underlying most RAG implementations — the agent queries the store with a relevant phrase and retrieves the most similar documents or facts.
The type of memory an agent uses determines what it can reliably know and act on. An agent without persistent memory cannot resume a task left incomplete in a prior session, cannot remember a client's preferences from a previous interaction, and cannot accumulate operational knowledge over time. These limitations are significant constraints on the complexity of tasks an agent can handle.
Conversely, an agent with poorly managed persistent memory may act on outdated information — quoting a contract rate that changed three months ago, referencing a contact who has left the organisation, or applying a policy that has been superseded. Memory quality directly affects output quality.
For Australian organisations, persistent memory stores that contain personal information attract obligations under the Privacy Act 1988. Subject access requests, correction rights and data retention limits under the Australian Privacy Principles apply regardless of whether the data is held in a traditional database or an AI agent's memory store.
In-context memory is managed by the orchestration layer, which constructs the context window for each inference call by assembling the system prompt, relevant retrieved content, conversation history (or a summarised version of it) and any tool outputs from the current session.
External persistent memory is implemented as a read/write interface to a database — typically a key-value store for structured state, a relational database for structured records, or a vector store for semantic retrieval. The agent calls a "write memory" tool to persist information and a "read memory" or "search memory" tool to retrieve it.
Episodic memory requires a logging and retrieval layer: agent interactions are structured, stored with metadata (timestamps, task types, outcomes), and retrieved using similarity search or filtered queries when the agent needs to recall analogous past situations.
Semantic memory is implemented via a RAG pipeline — embeddings are generated for stored content, and queries are encoded and matched against stored embeddings to surface the most relevant information.
A production agent may use all four types simultaneously: in-context memory for the current task, external memory for state across sessions, episodic memory for relevant precedents, and semantic memory for knowledge retrieval.
Memory architecture must be designed before implementation begins, not added retrospectively. The key decisions are: what information needs to persist beyond a single session, in what format, for how long, and who can access or modify it.
Scoping memory carefully is essential. Agents should write only what they genuinely need to retrieve later — promiscuous memory writes inflate storage costs, create data governance surface area and increase the risk of stale-data errors. A memory schema — defining what categories of information the agent is permitted to store — is the equivalent of a data dictionary for the agent's memory system.
Memory expiry is frequently overlooked. Information that was accurate when written becomes stale over time. Implement TTL (time-to-live) policies on memory records and build a scheduled review or re-validation process for long-lived state that does not naturally expire.
Edison AI's AI implementation team designs memory architecture as a distinct workstream in complex agent builds — specifying memory types, storage technologies, access patterns and governance rules before the agent's capabilities are built on top of them. Retrofitting memory governance onto a running system is significantly more costly than designing it in from the start.
For any agent deployment that spans multiple sessions or handles tasks with state — onboarding workflows, account management, project execution — define the memory architecture explicitly as part of the design phase. Specify the four memory types and which are required, design the memory schema, set data retention and expiry policies, and confirm that your data governance framework covers the agent's memory stores alongside your other data assets.
Edison AI designs and ships AI agents and workflow automation built around how your business actually runs.
AI agents do not have persistent memory by default — each inference call begins with a fresh context window. Agents maintain information across tasks by explicitly writing state to external storage (databases, files, vector stores) and retrieving it at the start of subsequent tasks. In-context memory holds information only for the duration of a single interaction.
The main types are in-context memory (information held in the active context window for the current task), external or persistent memory (information written to and retrieved from databases or files between sessions), episodic memory (records of specific past interactions, stored and retrievable), and semantic memory (general knowledge or facts stored in a vector database and retrieved by similarity search).
Persistent memory can accumulate outdated, incorrect or sensitive information over time. Agents that retrieve stale context may make decisions based on information that is no longer accurate. Memory stores containing personal information are subject to Australian Privacy Act obligations, including rights of access and correction. Memory must be scoped, reviewed and expired on a defined schedule.
Edison AI helps Australian businesses move from AI curiosity to practical implementation, with workflow design, team training and measurable outcomes. Tell us about your setup and we'll come back with a sequenced plan grounded in the same thinking you just read.
Article: How AI Agents Use Memory to Maintain Context Across Tasks
