What are the three types of memory in the context of AI agents?

The three types of memory mapped from human cognition are semantic memory (facts about a user, like knowledge learned in school), episodic memory (past agent actions or specific experiences), and procedural memory (instructions or system prompts for the agent).

How do agents update their memories?

Agents can update memories in two ways: 'hot path' where memories are updated before responding to the user, offering transparency but potentially increasing latency; and 'background' where updates happen later, reducing latency but being less immediate or transparent to the user.

How can memory be managed in an enterprise setting with multiple users?

Memory in an enterprise setting can be managed by namespacing. This involves associating memories with specific users, organizations, or the agent itself, allowing for tailored access control and preventing memory conflicts or privacy issues between different entities.

How can token usage be controlled when using memory in AI agents?

Controlling token usage involves strategically choosing where to apply memory updates, prioritizing high-leverage points such as when an LLM explicitly signals a lack of information or during user feedback. Thoughtful UX design also helps by allowing developers to intelligently place memory updates.

What tools are used to add semantic memory to an agent?

Semantic memory is added to an agent using tools like `create_manage_memory_tool` and `create_search_memory_tool` from the Langmem package. These tools allow the agent to update (add, edit, delete) and search for facts about users, respectively.

What is the role of 'f-shot prompting' in episodic memory for email triage?

F-shot prompting is used in episodic memory for classification tasks, such as email triage. It involves providing the LLM with examples of previous emails and their classifications (e.g., ignore, respond), allowing the agent to learn and apply these patterns to new emails.

Why is User Experience (UX) important for memory in AI applications?

UX is crucial for AI memory because it enables human-in-the-loop interaction, which improves performance, builds user confidence, and facilitates the collection of valuable user feedback. This feedback can then be used to grow and correct memory, such as storing a user's decision to ignore an email as an f-shot example.

How does procedural memory relate to prompt optimization?

Procedural memory in AI agents directly corresponds to the system prompts and instructions given to the agent. By dynamically updating these prompts based on user feedback or automated evaluation, the agent's instructions (and thus its behavior) are optimized over time, treating memory and prompt optimization as two sides of the same coin.

How are prompts updated based on feedback in procedural memory?

An LLM-based agent, often a Langraph agent, is used to reflect on trajectories (conversations) and associated feedback. It classifies which prompts need updating and then uses another LLM call to generate revised prompt instructions, which are then stored and used by the agent.

How can prompt hallucinations be addressed in AI agents with memory?

Addressing prompt hallucinations requires human-in-the-loop oversight, allowing users to correct erroneous agent behavior in real-time. Developers can also manually inspect and edit accumulated memories, and observability tools like Langsmith are essential for debugging and tracing the source of hallucinations.

What are some practical applications of Langchain's agent capabilities?

A practical application is a social media agent that monitors Langchain community contributions, drafts posts to highlight them, and allows for user interaction and feedback to refine its posting style and content, demonstrating how memory and dynamic prompt updates lead to a more effective system.

AI Dev 25 | Harrison Chase: Long Term Memory with LangGraph

DeepLearning.AI

Entertainment4 min read92 min video

Mar 27, 2025|2,823 views|38|3

Save to Pod

Key Moments

On this page

TL;DR

AI agents can use semantic, episodic, and procedural memory, managed via LangGraph, for personalized interactions.

Key Insights

AI agents can be endowed with long-term memory by categorizing it into semantic (facts), episodic (experiences), and procedural (skills) types.

LangGraph provides a framework for building agent and multi-agent systems, enabling the implementation of complex memory functionalities.

Memory can be updated in real-time ('hot path') or asynchronously ('background'), each with trade-offs in latency and transparency.

Procedural memory directly maps to agent instructions and system prompts; its optimization is akin to prompt engineering.

Episodic memory, often implemented using few-shot examples, is effective for classification tasks like email triage.

Semantic memory, analogous to RAG, stores factual information about users or the environment, retrievable via semantic search.

THE IMPORTANCE OF MEMORY IN AI AGENTS

AI agents, to be truly effective personal assistants, need to emulate human-like memory. Unlike static LLM interactions that require repetitive context provision, agents with memory can recall past interactions, learn user preferences, and adapt over time. This is crucial for moving agents from being merely 80% useful to achieving 95-97% effectiveness, enabling personalization and a more natural, less redundant user experience.

TYPES OF MEMORY MAPPED TO HUMAN COGNITION

The presentation categorizes AI memory into three types, drawing parallels with human memory. Semantic memory stores factual information, similar to facts learned in school. Episodic memory captures specific past experiences and events, like remembering a particular conversation. Procedural memory encompasses skills and behaviors, akin to learning how to ride a bike. These categories provide a conceptual framework for designing and implementing memory systems in AI agents.

IMPLEMENTING MEMORY TYPES IN AGENTS WITH LANGGRAPH

These memory types are mapped to agent functionalities. Semantic memory is represented by facts about the user, retrievable through techniques like semantic search. Episodic memory can be implemented using few-shot prompting with past agent actions as examples. Procedural memory aligns with system prompts and instructions that guide the agent's behavior and tone. LangGraph, an orchestration framework, is used to build these complex agent systems.

MEMORY UPDATE STRATEGIES: HOT PATH VS. BACKGROUND

Two primary strategies exist for updating memory: 'hot path' and 'background'. Hot path updates occur before the agent responds, allowing for immediate use of new information but potentially increasing latency and complexity for the agent. Background updates happen asynchronously after the interaction, avoiding latency issues and separating memory management but with less immediate user visibility. Both approaches have distinct advantages and are utilized in agent development.

SEMANTIC MEMORY: FACTS AND USER INFORMATION

Semantic memory stores factual data, such as user preferences or specific details about past interactions. Implemented using tools like `manage_memory` and `search_memory` from the LangMe package, it leverages an in-memory store initially, with production-ready options like PostgreSQL and Redis available. This memory is namespaced, allowing for user-specific or organization-specific storage, and is crucial for an agent to recall facts pertinent to its tasks.

EPISODIC MEMORY: PAST INTERACTIONS AND EXPERIENCES

Episodic memory, particularly useful for classification tasks like email triage, is implemented using few-shot examples. These examples, representing past interactions and their desired outcomes, are stored in a memory store and retrieved via semantic search. For instance, an agent can learn to classify emails as 'respond,' 'ignore,' or 'notify' based on curated historical examples, allowing it to generalize and handle new, similar situations effectively.

PROCEDURAL MEMORY: INSTRUCTIONS AND BEHAVIORS

Procedural memory directly translates to the agent's instructions and system prompts. This type of memory is crucial for defining the agent's core behaviors, tone, and specific task guidelines. By storing these instructions in a persistent store and fetching them at runtime, they can be dynamically updated based on user feedback or performance analysis. This approach blurs the line between memory and prompt optimization, allowing agents to adapt their fundamental operating procedures.

THE ROLE OF USER INTERFACE AND FEEDBACK LOOPS

A user-friendly interface is critical for gathering feedback, which in turn fuels memory updates and prompt optimization. Whether it's a simple 'ignore' button on an email, an option to edit generated content, or direct natural language instructions, user interaction provides invaluable data. This feedback loop allows agents to learn, adapt, and refine their responses, moving towards more sophisticated and personalized assistance, and it is essential for correcting potential hallucinations or suboptimal behaviors.

PERSISTENCE AND NAMESPACING FOR MEMORY MANAGEMENT

Memory persistence is managed through a 'store' that acts as a namespace key-value store supporting semantic search. This allows memories to be scoped to specific users, organizations, or agents. For enterprise settings, robust authentication mechanisms are layered on top of these namespaces to ensure data privacy and prevent proliferation of information between different agents or users, safeguarding sensitive data.

OBSERVABILITY AND DEBUGGING WITH LANGSMITH

Tools like LangSmith are essential for observing and debugging agent behavior, especially when memory is involved. They provide detailed traces of agent execution, including LLM calls, tool usage, and memory interactions. This transparency is vital for understanding why an agent behaves a certain way, identifying potential issues like hallucinations, and iterating on prompt design and memory management strategies. Bulk exports to data warehouses further aid in system-level observability.

ADVANCED CONSIDERATIONS: PRIORITIES AND MULTIMODALITY

While not explicitly covered in basic implementations, advanced memory systems can incorporate priority mechanisms, combining semantic similarity, importance scores, and timestamps to resolve conflicting memories. The system's current store primarily handles key-value lookups for multimodal data; semantic search over multimodal content is an area for future development. The underlying principles of memory management, however, remain applicable across different data types and complexity levels.

Mentioned in This Episode

●Software & Apps

●Companies

●Organizations

●Books

Common Questions

Langchain is a set of developer tools designed to make it easy to build applications on top of Large Language Models (LLMs). Langraph is an orchestration framework within Langchain specifically for building highly customizable agent and multi-agent systems, often serving as a follow-up to Langchain for more advanced agent development.

Topics

F-shot Prompting User Experience (UX)Token Usage Optimization

Mentioned in this video

softwareRedis

A database system mentioned as a production-ready persistence layer for long-term memory storage in Langraph.

softwareLangSmith

A debugger for observing agent trajectories and internal LLM traces, used in debugging and understanding agent behavior.

softwareRoBERTa

A natural language inference model mentioned as a possible tool to check for conflicting memories in a background process.

bookGenerative Agents

A paper from Stanford on simulating agents, which spent significant time on memory systems, including importance scores and temporal aspects for memory retrieval.

companySplunk

A system where observability data from Langsmith can be published for system-level monitoring.

toolLangChain