Auto Dream vs OpenClaw Dreaming: How AI Agents Consolidate Memory

Background: Noise Accumulation in Agent Memory

When an AI agent accumulates many notes, contradictory or outdated entries remain, causing ambiguous decisions. An example shows a virtual assistant recording both "Boss prefers plan A" and later "Boss prefers plan B" without deleting the old note, leading to confusion. Relative dates such as "yesterday" become ambiguous after a month, turning the memory into a noisy library.

Auto Memory – Claude Code’s Automatic Note‑Taking

Since February, Claude Code enables a default Auto Memory that records observations (frameworks, code style, project architecture) into a project‑isolated directory ~/.claude/projects/<project_name>/memory/. The directory contains: MEMORY.md – index file user.md – user information feedback.md – corrections or affirmations project.md – project progress reference.md – external resources

At the start of each new conversation Claude reads the first 200 lines of MEMORY.md and calls the other files on demand.

Auto Dream – Claude Code’s Four‑Step Consolidation

The system prompt declares the agent is "dreaming" – a reflective scan of its memory files. The four steps are:

Orient : read the entire memory directory to understand existing files and relationships.

Gather Signals : search past conversation logs for correction points, user‑requested facts, repeated items, and important decisions.

Consolidate : compare signals with current memory, merge duplicates, resolve contradictions, correct dates, and remove stale entries.

Prune & Index : delete redundant information and re‑index the cleaned memory.

Effectiveness: in a community case with 913 dialogues, Auto Dream reduced MEMORY.md from 280 lines to 142 lines in 8–9 minutes, fixing contradictions and updating outdated framework names.

Trigger conditions require both a 24‑hour interval since the last run and at least five new dialogue records. The process runs in a background sub‑agent, leaving the active conversation unaffected. Isolation ensures only memory files are writable; source code remains read‑only, and only one Auto Dream can run per project.

OpenClaw Dreaming – Three‑Stage Collaboration

OpenClaw’s Dreaming mirrors human sleep cycles: Light (shallow), Deep (core decision), and REM (pattern recognition). Each stage handles signals differently:

Light : organize recent short‑term signals, de‑duplicate, store candidates in DREAMS.md, never write to MEMORY.md.

Deep : rank candidates using a six‑dimensional weighted score plus reinforcement signals from Light and REM; only items passing all thresholds are promoted to MEMORY.md.

REM : extract patterns from recent short‑term traces, generate reflective summaries in DREAMS.md, and provide reinforcement signals for Deep ranking without writing to long‑term memory.

Adjustable parameters such as recencyHalfLifeDays and maxAgeDays fine‑tune memory aging. The mode is controlled via commands /dreaming core, /dreaming balanced, or /dreaming deep.

Claude Auto Dream vs OpenClaw Dreaming – Core Differences

Process : Claude uses a four‑step pipeline (Orient → Gather → Consolidate → Prune); OpenClaw follows three sleep‑inspired stages (Light → Deep → REM).

Trigger : Claude requires 24 h + 5 dialogues; OpenClaw relies on the dreaming.mode setting.

Promotion Mechanism : Claude directly edits MEMORY.md without explicit scoring; OpenClaw uses a six‑dimensional weighted score with threshold gates.

Signal System : Claude’s signals are implicit, derived from dialogue logs; OpenClaw’s are explicit, combining weighted scores and stage‑specific reinforcement.

Output Files : Claude updates MEMORY.md and sub‑files; OpenClaw writes a summary to DREAMS.md and promotes entries to MEMORY.md.

Adjustable Parameters : Claude offers limited knobs; OpenClaw exposes recencyHalfLifeDays, maxAgeDays, etc.

Isolation : Both enforce read‑only code access; only memory files are writable.

Concurrency Control : Claude allows only one Auto Dream per project; OpenClaw does not specify concurrency.

Four‑Layer Memory Architecture in Claude Code

CLAUDE.md : hand‑written directives (project standards, coding guidelines) with highest authority.

Auto Memory : automatically captured notes during work.

Session Memory : transient context for a single conversation; raw logs are kept as JSONL files for later review.

Auto Dream : background consolidation layer that periodically cleans and optimizes accumulated notes.

Cognitive‑Science Perspective

The design draws from human memory stages: sensory, short‑term, and long‑term memory, and the consolidation that occurs during non‑REM and REM sleep. Auto Memory corresponds to the hippocampal short‑term cache, while Auto Dream’s four steps emulate nightly replay, integration, and selective forgetting performed by the cortex. Unlike humans, AI retains the original JSONL dialogue logs untouched, enabling selective forgetting without losing auditability.

Implications for Agent Engineering

Traditional approaches (adding more context or RAG retrieval) cannot solve information overload. The new mechanisms shift the focus from memorizing everything to learning to forget selectively, providing sustainable memory health for enterprise‑grade agents. OpenClaw’s six‑dimensional scoring includes a “cross‑day consolidation” signal, favoring information that proves valuable over multiple days rather than short‑term spikes.

Current Status and Outlook

Both features are experimental and disabled by default. OpenClaw’s Dreaming already has over 100 contributors, indicating rapid maturation toward becoming a standard component of agent platforms.

Related Links

OpenClaw Release: https://github.com/openclaw/openclaw/releases/tag/v2026.4.5

OpenClaw Documentation: https://docs.openclaw.ai