What Is AI Agent Memory? 4 Types Explained (2026)

AI agent memory is the set of mechanisms by which an AI system stores, retrieves, and uses information across tasks, sessions, and time. Without memory, an AI agent starts from zero with each interaction. With it, the agent knows your past decisions, your current goals, and the history of every action it has taken on your behalf. The difference in output quality is not subtle.

The majority of AI tools sold today operate within a single session: they know what you told them in the current conversation and nothing more. When that session ends, the context disappears. A few tools, and an increasing number of agentic products, work differently. They carry information forward, retrieve it when relevant, and build an understanding of your work over time.

The technical term for how they do this is agent memory. The mechanics differ significantly across systems, and those differences matter when evaluating any AI product that claims to learn your context, personalize its outputs, or act on your behalf over extended periods.

What Are the Main Types of AI Agent Memory?

Researchers and engineers building AI agents describe four primary memory types: in-context memory, external memory, episodic memory, and semantic memory. Each plays a different role. Most capable AI agents combine at least two of them. The type of memory a product uses determines what it can actually know about you and what it can do with that knowledge.

The four types, in order of how they are usually implemented:

1. In-Context Memory (Working Memory)

In-context memory is the information the AI holds during a single session. It includes everything in the current conversation or task window: your instructions, the outputs produced so far, tool results, and any documents you have provided. Large language models access in-context memory instantly; it is loaded into the model at inference time.

The limit is the context window. GPT-4o, Claude, and Gemini each support different token limits, ranging from tens of thousands to hundreds of thousands of tokens. When a session ends, in-context memory is discarded. Nothing carries forward unless the system has another memory mechanism to store it.

Working memory is why an AI that knows everything about your task at minute one of a session may produce worse outputs at minute sixty: the context has grown, earlier content has been pushed further from the model's active attention, and the quality of retrieval degrades.

2. External Memory (Retrieval-Augmented Generation)

External memory is stored outside the model and retrieved when needed. The common implementation: a database stores documents, notes, conversation summaries, or factual records as vector embeddings. When the agent encounters a relevant query, it searches the database, retrieves the closest matches, and injects them into the current context before generating a response.

This is the mechanism behind "second brain" AI tools, knowledge management products, and long-horizon AI agents that need to recall information from weeks or months ago. Mem.ai, NotebookLM, and similar tools are built on external memory retrieval.

The trade-off is retrieval quality. External memory is only as good as what is stored, how it is indexed, and whether the retrieval step surfaces the right chunk at the right moment. Poorly chunked documents or weak embedding models produce confident-sounding outputs grounded in the wrong retrieved content.

3. Episodic Memory (Past Interactions)

Episodic memory is the record of what an agent has done before: which actions it took, what responses those actions generated, and what the outcome was. For a sales agent, this means knowing that it sent a connection request to a prospect on April 14th, that the prospect accepted, that a follow-up message was sent on April 18th, and that no reply came.

Episodic memory is what lets an agent pick up where it left off rather than starting fresh each session. Without it, every interaction is cold: the agent has no knowledge of prior conversations, prior actions, or prior results. With it, the agent can make decisions based on what actually happened rather than on general instructions.

For sales and outreach use cases, episodic memory is often the most valuable type. The personalization that matters to a buyer, "this message references our last conversation," requires the agent to have a reliable record of that last conversation.

4. Semantic Memory (Domain Knowledge and Facts)

Semantic memory is stored knowledge about the world, about a domain, or about a specific entity. It is the repository of general facts that the agent draws on when making decisions. For a sales agent, semantic memory includes information about target accounts: their industry, size, recent news, key personnel, and buying signals.

Unlike episodic memory, semantic memory is not tied to specific past events. It is knowledge that remains true across many interactions. The agent draws on it to frame a message correctly for a given industry, to avoid pitching a product the prospect's company is known to use a competitor for, or to reference a relevant event in an opening line.

Semantic memory is typically built from structured data sources, CRM records, and curated knowledge bases rather than from the agent's own interaction history.

How Do These Memory Types Work Together?

A well-designed AI agent combines multiple memory types to produce coherent, context-aware behavior across time.

Consider a sales agent handling a LinkedIn outreach sequence:

The agent holds the current task in in-context memory: the prospect's name, the draft message, the instructions for this type of outreach. When it needs to recall whether this prospect was contacted before, it retrieves from episodic memory. When it needs to know the prospect's company's recent funding announcement to personalize the message, it retrieves from semantic memory. If the user has a note about their preferred tone for messages to finance executives, external memory retrieval surfaces it.

Each memory type handles a different question. Together, they let the agent behave like a colleague who knows the situation rather than like a tool encountering each task cold.

Why Does Agent Memory Matter When Choosing an AI Tool?

Whether a tool has real memory, and what kind, determines whether it can personalize at scale. An AI product that restarts from scratch with each session can draft well but cannot learn. One with persistent episodic memory can track 500 prospects across a 90-day sequence without losing the thread of any of them.

Context engineering is the discipline of deciding what information an AI receives at each decision point. Memory is the mechanism that makes context engineering practical over time: without it, a user must manually re-inject relevant context at every session. With it, the agent pulls that context from storage.

For buyers evaluating AI products:

Tools with only in-context memory start fresh every session. They are useful for single-session tasks: drafting a document, analyzing a dataset, answering a question. They are not suitable for ongoing work that requires continuity.

Tools with external memory can recall stored information across sessions, but retrieval quality varies. Ask any vendor about their chunking strategy and embedding model, because the answer reveals how seriously they have thought about it.

Tools with episodic memory can track what they have done on your behalf across weeks and months. For outreach, research, and workflow agents, this is the capability that separates a capable agent from a sophisticated prompt interface.

Tools with semantic memory built into your domain or accounts can make better decisions without requiring you to explain the context every time. The difference shows up in output quality on nuanced tasks.

How Northlight Uses Agent Memory for Sales Outreach

Northlight is a personalized AI sales agent that handles LinkedIn and email outreach. The memory design is specific to the sales use case.

Episodic memory: Northlight knows the full interaction history with each prospect in your list. It knows whether you sent a connection request, whether they accepted, what messages were exchanged, and what the current stage of the conversation is. This is what lets it continue a sequence intelligently rather than restarting it.

Semantic memory: Northlight carries knowledge about your target accounts and criteria: the industries you are focused on, the personas you reach, the signals you use to prioritize. This informs which prospects get which type of outreach.

In-context memory: Each drafting session holds the current task, the prospect's context, and your voice instructions. This produces message drafts that match your style rather than generic templates.

The structural advantage is that Northlight runs through your real browser session rather than through a cloud integration. This means the episodic record is tied to your actual accounts and their actual state. What LinkedIn records as sent is what Northlight records as sent; the two cannot diverge.

Pro starts at $100/month ($80/month billed annually). Ultra is $200/month ($160/month annually). See plans.