7-langgraph.md•3.27 kB
# 🧠 LangChain's Second Strategy for Long-Term Memory
LangChain recently introduced a second approach to agent memory — inspired by **how humans organize memory**.
This design was shared by LangChain's CEO during a course with DeepLearning.AI, and it's structured around three types of memory:
**Semantic, Episodic, and Procedural**.
---
## 🧩 Types of Memory Modeled
### 1. Semantic Memory
- Stores **factual knowledge** about the world or the user
- Examples:
- “User prefers morning meetings”
- “John's favorite food is pizza”
- Stored in a **VectorDB** for semantic similarity retrieval
### 2. Episodic Memory
- Stores **past experiences** and events
- This includes full or summarized chat interactions, task logs, and personal anecdotes
- Helps the agent remember “what happened”
### 3. Procedural Memory
- Stores **instincts or learned behaviors**
- These are injected into the **system prompt** and define how the agent should behave
- Example:
- “Always greet the user warmly”
- “Maintain a formal tone with this user”
---
## ⚙️ How It Works — The Memory Pipeline
- Is it a fact? → Implement it in **Semantic Memory** (This is implemented within the tools that we provide to the agent.)
- Is it a past event or story? → Implement it in **Episodic Memory** (This includes the few-shot examples provided by the user and stored in the vector database. The user can also add more examples over time.)
- Is it behavioral guidance? → Implement it in **Procedural Memory**
LangChain uses a **multi_prompt_optimizer** LLM. When the user provides feedback to adjust the behavior of the two agents, this LLM decides which agent’s prompt should be updated and how to update it.
Then, when the user sends a new message:
1. The system **analyzes and routes the memory** appropriately
2. It **retrieves relevant past context** from the different memory types
3. It **builds a dynamic system prompt** with updated facts, experiences, and behavior rules
4. The LLM generates a response based on this enriched context
---
## 🔧 Tools Used by the Agent
To interact with these memory types, LangChain uses dedicated tools:
- `manage_memory_tool` – update or revise memory entries
- `search_memory_tool` – retrieve semantic or episodic data
- `writing_tool`, `calendar_tool`, `scheduling_tool` – perform tasks based on remembered facts
The LLM can call these tools **intelligently and autonomously** depending on the context.
It can:
- Take actions
- Decline to act
- Or ask the user for clarification
---
## 🗺️ Architecture Overview
This architecture supports:
- Adaptive memory routing
- Real-time context injection
- Human-like memory segmentation
It's one of the most **comprehensive and flexible** memory strategies in production-ready LLM systems today.
---
## 👨🏫 Let's See It in Action
Now let's step into the code.
We'll explore:
- How memories are created and stored
- How LangChain routes and retrieves them
- And how the agent uses this layered memory to behave intelligently in real time
Let's dive in.