Part 4¶
Overview¶
Memory is what separates stateless chat from stateful agents. Without memory, agents cannot learn, maintain context, or improve over time.
This section covers:
- How memory works in agentic systems
- Short-term vs long-term memory
- Memory retrieval and compression
- Episodic and procedural memory
Memory Types at a Glance¶
| Type | Duration | Use Case | Storage |
|---|---|---|---|
| Context Window | This turn | Current task | LLM |
| Episodic | Hours/days | What happened | Vector DB |
| Semantic | Persistent | Knowledge | KB/Graph |
| Procedural | Persistent | How to do things | Learned patterns |
Why Memory Matters¶
- Context: Keep track of ongoing work
- Learning: Remember past successes and failures
- Grounding: Anchor to specific past events
- Efficiency: Don't recompute what's known
-
Chapter Map¶
- Memory Fundamentals
- Short Term Memory
- Long Term Memory
- Vector Stores & Retrieval
- Memory Compression
- Episodic & Procedural
Key Challenge¶
The Context Window Dilemma:
- LLMs have finite context windows (100K-1M tokens)
- Agents generate lots of tokens (actions, reflections, observations)
- Must choose what to keep and what to forget
-
Next: Start with Memory Fundamentals¶
-
Last Updated: August 9, 2026