Memory Dreaming

Signal-aware memory consolidation in three phases. Instead of processing logs by date order, this skill reads memory/.dreams/short-term-recall.json (maintained automatically by OpenClaw's memory-core) to find what's been most frequently recalled — then prioritizes consolidating those entries into long-term topic files.

Memory Architecture Assumed

This skill expects a two-layer memory layout:

• Daily logs (memory/YYYY-MM-DD*.md) — raw session notes, never deleted
• L2 topic files (memory/<topic>.md) — curated, durable knowledge on a specific subject (e.g. memory/clash-verge.md, memory/business.md). Created manually or by the dream agent when new topics emerge.
• Index (MEMORY.md) — high-level status overview with pointers to L2 files

If you are starting fresh, create at least MEMORY.md and memory/dream-log.md before the first dream run. L2 files are created on demand.

Quick Start

Manual dream (run once)

Tell your agent:

"Run a memory dream consolidation. Follow the protocol in <SKILL_PATH>/references/dream-protocol.md. Workspace root: <YOUR_WORKSPACE_PATH>."

Or simply: "爆爆做个梦" / "run a dream consolidation" — if this skill is loaded, the agent will know what to do.

Automated daily dream (cron)

Set up an isolated agentTurn cron that runs while your human is asleep:

{
  "name": "daily-dream",
  "schedule": { "kind": "cron", "expr": "0 7 * * *", "tz": "<YOUR_TIMEZONE>" },
  "sessionTarget": "isolated",
  "payload": {
    "kind": "agentTurn",
    "timeoutSeconds": 900,
    "message": "Run a memory dream consolidation. Find and read references/dream-protocol.md inside the signal-dreaming skill (check your skillDirs; common locations: skills/signal-dreaming/ or ~/.openclaw/skills/signal-dreaming/). Workspace root: <YOUR_WORKSPACE_PATH>. End your final response with a one-line dream summary — the cron delivery mechanism will auto-announce it."
  },
  "delivery": { "mode": "announce", "channel": "<CHANNEL_TYPE>", "to": "<CHANNEL_ID>" }
}

Adjust expr and tz to match when your human sleeps.

How the Signal Works

OpenClaw's memory-core maintains memory/.dreams/short-term-recall.json automatically. Every memory_search call updates this file with:

• recallCount — how many times a snippet was returned in results
• totalScore — weighted composite (relevance × frequency × query diversity × recency)
• queryHashes — distinct queries that surfaced this snippet (diversity signal)
• recallDays — days the snippet was active

The dream protocol reads totalScore to rank candidates. Snippets with totalScore > 0.8 from multiple distinct queries are strong promotion candidates.

Three-Phase Safety Model

Phase 1 is always read-only. An error in Sense never corrupts files.

Quality Gates

The full protocol includes a read-only planning checkpoint before writes:

• Topic identity guard — do not merge legacy/current projects just because names are similar; split active, legacy, and index files when durable IDs, hosts, owners, or lifecycle differ.
• Lifecycle guard — classify candidates as active / waiting / done / archived / closed / snowed; closed or snowed work must not reappear as active TODOs unless the human reopens it.
• Secret propagation guard — never promote API keys, tokens, passwords, private keys, signed URLs, or credential-bearing command lines from daily logs into L2, MEMORY.md, or dream-log.md.
• Backup guard — existing L2 files are copied outside memory/ to <WORKSPACE_ROOT>/.backup/memory-dreams/YYYYMMDD-HHMM/ with a .bak suffix before modification; MEMORY.md is backed up there too before rewrite.
• Post-write audit — check size, Markdown structure, current-vs-legacy separation, and obvious credential patterns before finalizing. Use references/dream-audit.sh for the common lightweight sanity checks; it is not full DLP.

Compatibility with OpenClaw 2026.4.15+

OpenClaw 2026.4.15 changed the built-in memory-core Dreaming default from inline to separate mode. This is a different system from this skill:

These two systems are independent. You can use this skill whether or not built-in Dreaming is enabled.

If you use both:

• Keep built-in Dreaming in separate mode (default in 2026.4.15+). This keeps memory/YYYY-MM-DD*.md clean so this skill sees only real session notes.
• If you're on an older version with inline mode, the protocol will skip ## Light Sleep / ## REM Sleep phase blocks when scanning daily logs.

Key Rules

• Never move or delete daily logs (memory/YYYY-MM-DD*.md) — archiving breaks memory_search indexing
• dream-log.md is Markdown — append text directly, do not write JSON
• Never copy secrets into curated memory — omit/redact and alert instead
• Keep lifecycle state sticky — closed/archived/snowed lines stay non-active unless explicitly reopened
• Back up MEMORY.md to <WORKSPACE_ROOT>/.backup/memory-dreams/YYYYMMDD-HHMM/MEMORY.md.bak before rewriting it
• Back up touched L2 files outside memory/, with .bak suffix, before editing existing topic files
• MEMORY.md target: ≤ 8KB

Full Protocol

See references/dream-protocol.md for the complete three-phase workflow, scoring thresholds, dream-log format, and safety mechanisms.