Claude Code: Memory Consolidation (/dream)

It’s easy to get Claude to follow a set of instructions. It’s much harder to get it to remember the nuance of a 2,000-line codebase across twenty different sessions.

The default behavior of most AI agents is "read the transcript." But transcripts are messy. They contain failed attempts, backtracking, and irrelevant chatter. When a session starts, the agent spends 10 minutes (and 50,000 tokens) just trying to figure what the current state of the world is.

The Solution: Memory Consolidation

Instead of letting transcripts accumulate forever, we use a process called dreaming. At the end of a session, we ask Claude to synthesize the delta — the things that actually matter — into a set of durable memory files.

The next session doesn't read the old transcripts. It reads the Memories.

The Flow

1. Work: Do your regular coding session.
2. Dream: Run /dream (the skill below).
3. Persist: Claude updates memory/*.md and memory/MEMORY.md.
4. Wake: The next session starts with a compact, accurate index of everything that has happened.

Get the skill

Save the block below as .agents/skills/dream/SKILL.md in your project, then invoke it with /dream at the end of any session.

// Skill not found: undefined

The verdict

Run /dream at the end of every long session. It takes 30 seconds and it compounds — each session makes the next one faster. Paired with a solid MEMORY.md structure, it's the single highest-leverage habit I've found for working with Claude Code on long-running projects.