This video explains the difference between a chatbot that forgets you every session and an agent that learns. Just like you have short-term recall, stored facts, learned skills, and past experience, agents need the same. Watch this video for the four types, starting simple and building to the one where learning actually happens.
What this means for you
Start with what an agent sees right now. Working memory is the agent's context window: everything it can access in this session. Like RAM on a computer, it is fast and immediate, but temporary. When the session ends, it is gone. It also has a size limit. Even the biggest context windows top out.
The second type is semantic memory. This is where the agent stores facts, rules, documentation, things it needs to know in general. In production systems, this is often just Markdown files in a folder. The agent loads these at the start of each session so it always knows the ground truth. Without semantic memory, an agent would repeat the same mistakes because it has no persistent knowledge to draw from.
Third is procedural memory. This is how an agent knows how to do things. An agent skill is a folder with a Markdown file that describes what it does and step-by-step instructions. The clever part: an agent does not load all its skills into working memory at once (which wastes context). It keeps a lightweight index of what skills exist, then loads the full instructions only when it needs them. A skill might describe running a code review or building a presentation, and the agent pulls in extra files and templates only as it executes.
Fourth is episodic memory. This is the agent's record of what happened in past interactions, what worked, what it learned. The naive version is just to save every conversation and search through them. But production systems do something smarter. They distil and compress the experience. Rather than storing a full 45 minute debugging transcript, the agent might note: "the auth module issue was in the middleware layer." That is something useful to remember.
Not every agent needs all four. A simple thermostat or routing bot might only need working memory. A password reset agent needs working memory and procedural memory. A coding agent needs all four because it carries project knowledge across sessions and remembers what failed before.
Memory is what separates an agent from a chatbot. A chatbot gives a response and forgets. An agent gives a response shaped by persistent knowledge and accumulated experience.
Take a support team with a ticket queue. Early on, an agent might answer the same question three different ways because it has no persistent knowledge. With semantic memory loaded (the support playbook, company policies, product documentation), it answers consistently. With episodic memory working, the agent notes that tickets tagged "billing" often need escalation to finance. Next time it sees a similar ticket, it flags it earlier. Over weeks the agent gets better because it remembers what happens.
Try this
Sketch out an agent you want to build or use on your team. List what it needs to know (semantic memory), what skills it has to perform, and what mistakes you want it not to repeat. That gap between "what it knows now" and "what it needs to remember" is where memory matters.
Common questions about AI agent memory
How does AI agent memory work?
AI agent memory works through four types that build on each other, so the agent gives responses shaped by persistent knowledge and accumulated experience instead of forgetting everything when the session ends. Working memory is the context window for the current session, fast but temporary. Semantic memory is stored facts, rules, and documentation. Procedural memory is the skills that tell it how to do things. Episodic memory is a record of what happened in past interactions.
Take a support team with a ticket queue. Semantic memory gives it the playbook and policies to answer consistently, while episodic memory lets it notice that billing tickets often need escalation and flag them earlier.
What are the types of AI agent memory?
There are four types, each doing a different job:
- Working memory: the context window, everything the agent can see right now in this session, which disappears when the session ends.
- Semantic memory: facts, rules, and documentation the agent loads at the start of each session, often just Markdown files in a folder.
- Procedural memory: the agent's skills, the step-by-step instructions for how to do things.
- Episodic memory: its distilled record of past interactions and what it learned.
A support team's ticket bot might lean on semantic memory for the policies it loads each session and episodic memory for the lessons it picks up from past tickets. Not every agent needs all four, so match the types to the job rather than building everything.
Do AI agents remember previous conversations?
They can, through episodic memory, which is the agent's record of what happened in past interactions and what worked. The naive version saves every conversation and searches through it, but better systems distil and compress the experience, for example noting "the auth module issue was in the middleware layer" rather than storing a full 45 minute transcript. This is what separates an agent from a chatbot. On a support desk, a chatbot answers a ticket and forgets, while an agent improves over time because it remembers what billing tickets usually need.
What is the difference between short-term and long-term memory in AI agents?
Short-term memory is the agent's working memory: the context window, fast and immediate like RAM on a computer but temporary and size-limited, so it is gone when the session ends. Long-term memory covers the three types that persist across sessions: semantic memory for facts and documentation, procedural memory for skills, and episodic memory for lessons from past interactions. A simple routing bot might only need short-term working memory, while a coding agent needs all four because it carries project knowledge across sessions.