Memory architectures

You now have the pieces. This lesson studies three well-known architectures that combine them, so you can borrow the ideas that fit your problem.

Watch

Letta: Building Stateful AI Agents with Memory and Sleep-Time ComputeArize AI · 37 min

1. Tiered memory: MemGPT and Letta

MemGPT borrows from operating systems. The context window is like RAM: fast and small. External stores are like disk: large and slower. The agent itself gets tools to move information between tiers, which the paper calls virtual context management.

Takeaways to borrow:

• Keep a small always-in-context block for core facts (who the user is, the agent's persona and rules).
• Put the long tail in searchable external storage.
• Let the agent edit its own memory through explicit tools, so changes are visible and auditable.

The project is now maintained as Letta, with docs and open-source code.

2. Memory stream with reflection: Generative Agents

This paper stores every observation as a natural-language record in a memory stream. Retrieval scores each record on recency, importance and relevance, then injects the top ones. Periodically the agent reflects: it asks itself higher-level questions over recent memories and stores the conclusions as new memories.

Takeaways to borrow:

• A scoring function beats raw similarity. Start with score = relevance + recency + importance and tune weights on your own questions.
• Reflection turns scattered episodes into durable semantic knowledge ("the user always postpones Friday tasks").

3. Temporal knowledge graph: Zep and Graphiti

Zep stores memory in a graph built by the open-source Graphiti engine. Per the paper, it keeps raw episodes, extracted entities and relationships, and community-level summaries, and it tracks when facts were true separately from when the system learned them.

Takeaways to borrow:

• Facts change. Store validity intervals instead of overwriting, so the agent can answer both "where does she live?" and "where did she live in 2023?".
• Relationship questions ("who did I discuss this with?") suit a graph. This connects directly to the GraphRAG course.
• The paper's benchmark numbers come from its authors, so treat them as a reason to evaluate on your own data.

Resources

Check yourself

1. Add a scoring function to your lesson 4 retrieval and compare answers with plain similarity on 10 questions.
2. Implement a reflection job that reads the last week of memories and writes 3 to 5 higher-level facts.
3. Add valid_from and valid_until handling so superseded facts stay queryable as history.