FIND-20260325-025 — Claude Code Auto-dream: automated agent memory consolidation (REM-like background sub-agent)

Igor Kudryk (@fancylancer3991) shared a visualization of Claude Code's new 'Auto-dream' feature — an unreleased background sub-agent (server-side flag: tengu_onyx_plover) that periodically consolidates agent memory files across sessions. It mirrors human REM sleep: 4 phases — Orient (scan MEMORY.md + topic files), Gather Signal (grep transcripts narrowly for corrections, patterns, decisions), Consolidate (merge into topic files, convert relative to absolute dates, delete contradicted facts), Prune and Index (keep MEMORY.md under 200 lines). Triggered automatically when ≥24h elapsed AND ≥5 sessions accumulated since last run. Also triggerable manually via /dream command. Discovered by community analysis of /memory logs — not officially announced by Anthropic. Linked to UC Berkeley 'Sleep-time Compute' paper (April 2025) showing ~5x reduction in test-time compute via offline pre-processing.

Directly relevant to the ADLC pipeline: (1) Veille agent already uses a file-based MEMORY.md system at ~/.claude/agent-memory/veille/ — Auto-dream would automate what is currently done manually at end of each session. (2) ADLC pipeline agents write to ~/.claude/agent-memory/pipeline/state.md — Auto-dream could consolidate and prune this autonomously. (3) The 200-line MEMORY.md constraint is already enforced in veille agent memory — confirms current architecture is correct. (4) Absolute date conversion rule (relative → absolute in memories) is already practiced. (5) When Auto-dream rolls out broadly, ODS agents will benefit without code changes — memory quality will improve automatically over time. (6) The /dream command could be manually triggered between ADLC pipeline cycles to pre-consolidate memory before a major sprint.

Claude Code Auto-dream: automated agent memory consolidation (REM-like background sub-agent)

Source

ODS Impact

Security Review

Tags