Claude Code Just Dropped Memory 2.0

Short summary

Anthropic (Claude / Cloud Code) quietly released AutoDream — an experimental background sub‑agent that periodically consolidates, prunes, and compacts Claude / Cloud Code project memory files (MD files) so long‑term memory is cleaner, more durable, and less bloated.

What AutoDream does (features / behavior)

• Runs a background sub‑agent (a “dream”) that:
  • Gathers recent session info and user feedback.
  • Reads current memory files (memory.mmd and other .md context files).
  • Runs a reflective consolidation/pruning pass (a “dream prompt”) to synthesize and compress important context.
  • Stores updated, compacted memory files (indexes / one‑line descriptors rather than full dumps).
• Operates automatically in the background as a cleanup/organization system so new sessions feel “sharp” rather than fuzzy.
• Only touches memory (.md) files — it does not modify code or scripts.
• Adapts over time, learning what context is important for a given user or project.

“Sleep‑like” checkpoints — periodic consolidation improves long‑term storage and context freshness. (analogy referenced from Anthony’s tweet)

Benefits / rationale

• Less repetition between sessions (fewer things you must re‑explain).
• Reduced bloat by pruning unnecessary tokens and irrelevant context.
• Faster orientation and better recall in new sessions because important facts are easier to surface.
• Periodic consolidation acts like sleep: it checkpoints and refreshes long‑term context.

How it’s triggered / run

• Two main ways to run:
  • Manual: enable AutoDream via the /memory UI or invoke /dream (or use natural language like “run your autodream”). Note: CLI/skill labels may sometimes show “unknown skill” while the background dream still starts.
  • Automatic: likely scheduled by time (e.g., every N hours) or by session count (e.g., after X sessions). This scheduling behavior is inferred from community discussion, not from complete official docs.
• UI status indicators typically include: running/dreaming, never ran, last ran timestamp, or idle.
• AutoDream is a global toggle but operates against per‑project memory files.

Typical workflow / implementation steps (inferred)

1. Gather recent session metadata and messages.
2. Read existing memory files.
3. Load data into the sub‑agent and run a “dream” prompt to:
   • Synthesize durable memories,
   • Keep memory under line/token limits,
   • Link to files with one‑line descriptions.
4. Consolidate and prune redundant or irrelevant entries.
5. Write back updated, compacted memory files.
6. Repeat periodically.

Hands‑on examples from video

• Demo 1: AutoDream ran for ~10 minutes, reviewed 13 sessions, and improved/updated 5 MD files.
• Demo 2: AutoDream was enabled globally but “never ran”; invoking /dream reviewed 285 sessions and updated 3 MD files in ≈8 minutes.
• After a run you can inspect exactly what was added or removed and restore or further edit entries.

Limitations / caveats

• Experimental feature with a rolling rollout; behavior and timings are partly inferred from community tests (no complete official docs shown).
• Some UI/skill invocation inconsistencies reported (for example, “unknown skill” messages while the background process runs).
• The exact “dream” prompt and consolidation heuristics have not been publicly confirmed.

Practical guide / quick how‑to

1. Open the Cloud Code interface and go to /memory.
2. Open “edit cloud memory files” → toggle AutoDream on (hover & press Enter if the UI shows it as off).
3. Manually start a consolidation with /dream or by asking in natural language.
4. Check the status line for “dreaming” and open background tasks to view progress and changes.
5. View updated MD files and restore or refine entries if needed.

Sources / main speakers

• Video narrator / demo creator (unnamed YouTuber presenting demos and commentary).
• Anthropic / Claude / Cloud Code (product providing AutoDream).
• Anthony (tweet referenced for the sleep analogy).
• Community discussion on X (Twitter) and Reddit (used to infer triggers/behavior).