Claude Code can actually build long-term memory using Dreams

Memory is usually the biggest bottleneck for any LLM agent. Most people just jam more tokens into the context window and hope for the best, but that's a losing battle once the conversation hits a certain length. I've been experimenting with Dreams to handle this, and it's a much cleaner way to let an agent "remember" things across sessions without the typical context bloat.

The core idea here isn't just storing logs, but using a synthesis process—the "Dream"—where the agent reflects on its recent interactions and compresses them into core facts, preferences, or learned behaviors. Instead of searching through a massive database of raw chat history, the agent queries these synthesized memories.

If you want to set this up for your own AI workflow, you can implement a simple loop where the agent periodically summarizes its state. Here is a basic conceptual implementation of how you might structure the memory update:

{
  "memory_update": {
    "timestamp": "2023-10-27T10:00:00Z",
    "source_interaction_id": "session_456",
    "synthesized_fact": "User prefers Python over TypeScript for data processing scripts and dislikes verbose documentation.",
    "confidence_score": 0.95,
    "category": "user_preference"
  }
}

To make this a real-world deployment, you need a three-step pipeline:

1. Observation: The agent tracks key events or contradictions during a live session.
2. Dreaming: At the end of a session or a specific trigger, a separate LLM call processes these observations. It asks, "What did I learn about the user or the project that is worth keeping forever?"
3. Integration: These distilled insights are stored in a vector database or a simple JSON profile that gets injected into the system prompt of the next session.

This is a massive leap over standard RAG (Retrieval-Augmented Generation). RAG is great for finding a needle in a haystack, but "Dreaming" is about building a mental model of the user. When the agent starts the next session, it doesn't just have access to old documents; it has a refined understanding of your specific needs.

For those doing a deep dive into agentic memory, the trick is in the filtering. If you save everything, you're just back to square one with a bloated context. The "Dream" phase must be aggressive about discarding noise and only keeping high-signal information. This transforms the agent from a stateless tool into something that actually evolves as you use it.