How Claude’s Auto Dream Cleans Up AI Memory While You Code

Anthropic recently rolled out a gray‑scale feature called Auto Dream in Claude Code. Users can enable it by entering the /memory command and checking for the auto‑dream option. The feature activates only when two conditions are met: at least 24 hours have passed since the last memory sweep and the current project has accumulated a minimum of five dialogue records.

When triggered, Claude spawns a background sub‑agent that processes the project’s memory without interrupting the active conversation. The underlying memory system, Auto Memory , automatically records useful information during interactions and stores it under ~/.claude/projects/<em>project‑name</em>/memory/. The directory is organized by project and contains four main document types: user (personal info), feedback (corrections or affirmations), project (progress, decisions, background), and optionally reference (external resources). An index file MEMORY.md lists these files and is read (first 200 lines) at the start of each new dialogue.

As projects grow, memory files become noisy: outdated notes, contradictory entries, and relative dates (e.g., “yesterday”) lose meaning over time, turning the memory store into a “noise library” that confuses Claude. Auto Dream was created to address this problem.

The Auto Dream workflow consists of four steps:

Orient : Claude scans the entire memory directory to understand the current file landscape.

Collect signals : It searches recent dialogues for correction points, repeated instructions, important decisions, and other signals.

Consolidate : Signals are compared against existing notes; duplicate information is merged, contradictions are resolved, dates are normalized, and obsolete entries are removed.

Prune and index : Redundant or useless data is deleted, and the index file is regenerated.

In a small test, the process took 1 minute 19 seconds and reduced a 280‑line MEMORY.md to 142 lines, eliminating contradictory API error records and updating stale framework names. In a larger community example with 913 dialogues, Auto Dream required 8–9 minutes to clean up the memory.

Claude’s memory architecture now has four layers:

Layer 1 – CLAUDE.md: user‑written directives and project standards.

Layer 2 – Auto Memory: automatically generated notes during work.

Layer 3 – Session Memory: short‑term context for a single conversation.

Layer 4 – Auto Dream: periodic consolidation of Auto Memory.

Additionally, raw dialogue logs are stored locally as JSONL files; Auto Dream can read these logs to extract valuable information without altering the original records, preserving the ability to revisit past conversations.

The design mirrors human memory consolidation during sleep. Human memory progresses from sensory (seconds) to short‑term (15‑30 seconds) to long‑term (potentially unlimited) storage, with consolidation occurring mainly during non‑REM and REM sleep stages. During REM, the brain selectively reinforces important connections and discards irrelevant details—a process Claude emulates through Auto Dream’s selective forgetting and retention.

Thus, Auto Dream not only keeps Claude’s knowledge base tidy but also gives the AI a rudimentary sense of time, allowing it to reflect on past interactions, prune noise, and retain useful insights, much like the human brain does while dreaming.