Imagine a streaming service recommendation system that learns your preferences over time, getting better with every interaction. This is long term memory at work. It allows AI systems to learn, adapt, and improve across sessions, not just a single conversation. In this video, we'll explore the different types of long term memory including episodic, semantic, procedural, and shared memory. You'll learn how each one serves a different purpose and how together they enable smart AI systems that truly remember and grow. So what exactly is long term memory? Long term memory is persistent storage that remains accessible across multiple sessions. Unlike short term memory which is scoped to the current session, long term memory persists after the session ends so it can be reused in future interactions. An agent with long term memory can remember past experiences and use them to make better decisions in the future. It can recognize patterns, recall preferences, and build on previous interactions because the memories will still be there tomorrow, next week, or even months later. Think about the difference this makes. Without long term memory, every session starts from scratch. The agent has no history, no context, no accumulated knowledge. With long term memory, the agent becomes something more. It transforms from a stateless responder into an intelligent system that genuinely learns and adapts. There are several types of long term memory including episodic, semantic, procedural, and shared. Each serves a different purpose. Let's explore them one by one starting with episodic memory. Episodic memory is the agent's record of specific events and interactions. Think of it like a journal. It captures what happened, when it happened, and who was involved. Each memory includes metadata like timestamps and participants, giving the agent context about past experiences. For example, consider a customer service agent. With episodic memory, it can recall that this particular user had a billing issue three weeks ago, spoke with an agent named Jenny, and a credit was applied to her account to resolve the issue. There are two common ways to implement episodic memory. The first is conversational memory, storing complete conversation transcripts with speaker turns. This maintains full coherence and context letting the agent refer back to exactly what was said. The other type of episodic memory is summarization memory. This is where the conversations get compressed to preserve key insights without storing every word. This is more efficient for long histories since the agent can quickly retrieve the gist of past conversations without processing thousands of messages. Both approaches have their place. Conversational memory gives you precision. Summarization memory gives you efficiency. Many systems use a combination of the two by storing full transcripts for recent interactions and summaries for older ones. The second type of long term memory is semantic memory. While episodic memory is about specific events, semantic memory is about general knowledge and facts. It's what the agent knows about the world like concepts, relationships, and information that applies broadly. For example, an AI tutor might store facts like photosynthesis converts sunlight into energy or Python is a programming language. These aren't tied to a specific conversation or event. They're general knowledge that the agent can draw on in any context. Semantic memory builds a knowledge base that applies across all users in all sessions. When a user asks a factual question, the agent can retrieve relevant knowledge from its semantic memory and provide an accurate answer. This is the most common use case for semantic memory. Factual question answering and knowledge based reasoning. It's important to note that semantic caching and semantic memory are not the same thing. Semantic caching is a type of short term memory for reusing recent responses efficiently. Semantic memory is about permanently storing learned facts and knowledge, not temporarily caching responses. The third type of long term memory is procedural memory. This is about knowing how to do things. Think of it like muscle memory for an AI. Humans have procedural memory too. We don't consciously think about how to type on a keyboard or ride a bike. We do it automatically. That knowledge of how to perform tasks is procedural. For an AI agent, procedural memory captures workflows, best practices, and learned strategies. Think of a customer service agent that handles billing issues. Over time, it learns an effective process to first verify the account, then check recent transactions, then either resolve the issue or escalate to a human. That sequence of actions is stored in procedural memory. Instead of approaching every task from scratch, the agent draws on learned procedures that have worked before. Use cases for procedural memory include any scenario where the agent needs to execute multi step workflows like troubleshooting processes, onboarding sequences, or data analysis pipelines. Procedural memory makes the agent better at its job with experience. The fourth and final type of long term memory is shared memory. Unlike the others which are typically isolated to individual users, shared memory is accessible across multiple users, multiple agents, or multiple sessions. It enables collaborative intelligence. Think about a customer service team. Each agent may have their own episodic memory of interactions they've handled. But a shared knowledge base that contains product information, common solutions, and company policies is accessible to everyone. When one agent learns something useful, that knowledge can benefit the whole team. Shared memory is also useful in multi agent systems. If we have multiple AI agents working together on a task, they need a way to coordinate. Shared memory provides a common space where agents can store information for others to access. But this requires careful namespace design. We need to think about how to organize shared information so agents can find what they need. In practice, this often means treating shared as a metadata tag, allowing any type of memory to be promoted to a collaborative workspace while maintaining its original context. We also need access controls to balance collaboration with privacy. Not everything should be shared. Some memories are personal to a user or agent. While we've looked at these four types of memory individually, they rarely operate in isolation. In an adaptive AI system, these four memory types work together just like our own brains do. To understand how it works, let's look at a single interaction with a corporate travel assistant bot. Imagine a consultant named Sarah asks the AI agent to book a trip to the London office for the big product launch next week. Instead of just searching for flights, the agent can utilize its entire long term memory system. First, it hits the semantic memory. The agent retrieves general facts. It knows where the London office is, what the local currency is, and that travelers need a specific plug adapter for UK outlets. Next, it consults the shared memory. It checks the company's global travel policy and sees that for product launches, employees are allowed to book premium economy and must stay at one of three partner hotels. Then it checks the episodic memory. It looks at Sarah's personal history. It remembers that on her last trip to London, she rated a specific hotel one star because of construction noise. The system adapts by automatically filtering that hotel out of the options. Finally, it uses its procedural memory. The agent follows the standard operating procedure for corporate bookings. It knows the multi step process, fetch options, get Sarah's pick, send the receipt to the accounting software, and add the trip to her calendar. By combining these four types of long term memory, the agent becomes an adaptive AI system. It uses the past and the general to refine the how within the rules of the group to provide an intelligent solution. Our example showed an agent using all four types of memory, but that's not always necessary. For instance, an agent could easily only need semantic memory to serve as a diagnostic advisor that identifies likely causes for why a vehicle won't start based on specific symptoms. Or another could only use procedural memory to guide a user through the step by step process of replacing a faulty kitchen faucet. We demonstrated the everything at once scenario here simply to give you a complete mental map of the possibilities. Allowing you to scale back and pick only the specific tools your unique project requires. Great job. In this lesson, we explored long term memory and the four types that make AI systems truly intelligent. Episodic memory records specific events and interactions, the agent's diary of past experiences. Semantic memory stores general knowledge and facts that apply across all contexts. Procedural memory captures how to perform tasks, the workflows and strategies that make agents more effective over time. And shared memory enables collaboration, letting multiple users or agents access common knowledge. Together, these memory types create a comprehensive system. An agent might use episodic memory to recall a user's past issues, semantic memory to retrieve relevant facts, procedural memory to follow an effective resolution process, and shared memory to access team wide knowledge.