signal-dreaming

Safe single-owner staged memory consolidation with autonomous bounded index compaction.

Consolidate daily logs into L2 topic files and a compact MEMORY.md index. Preserve the proven v1.3.1 Sense → Consolidate → Settle model while adding deterministic P0 safety guards.

Release candidate: 3.0.0-rc.3.

Ownership contract

Use exactly one memory consolidation writer.

• Require built-in OpenClaw Dreaming to be disabled before any write.
• Refuse write mode when more than one enabled signal-dreaming cron is visible.
• Treat missing CLI fields, schema drift, unsafe paths, or ambiguous ownership as read-only failures.
• Never modify OpenClaw configuration, cron, Gateway, another agent, or memory-core internals.
• Do not read private plugin state, SQLite, recall locks, DREAMS.md, memory/dreaming/**, or memory/.dreams/**.

Collect three fixed read-only public CLI outputs into a private temporary directory, then run preflight:

bash

openclaw --version > <EVIDENCE_DIR>/version.txt
openclaw config get plugins.entries.memory-core.config.dreaming.enabled --json > <EVIDENCE_DIR>/dreaming.json
openclaw cron list --all --json > <EVIDENCE_DIR>/cron.json
node <SKILL_DIR>/scripts/preflight.mjs <WORKSPACE_ROOT> \
  --version-file <EVIDENCE_DIR>/version.txt \
  --dreaming-file <EVIDENCE_DIR>/dreaming.json \
  --cron-file <EVIDENCE_DIR>/cron.json

If any collection command fails, stop. Keep evidence local; do not paste cron payloads into chat. Create the evidence directory with owner-only permissions and remove it after preflight. Use the raw command outputs; do not hand-author alternative evidence schemas.

Use --read-only only for diagnosis when built-in Dreaming is enabled. A read-only pass does not authorize writes. Zero scheduled writers is acceptable for an explicitly invoked manual run and produces a warning. Every scheduled invocation must add --scheduled; it then fails unless exactly one enabled writer is visible.

Runtime support

• Minimum and tested OpenClaw: 2026.7.1-2.
• Minimum and tested Node.js: 22.23.1.
• Tested platform for this release candidate: macOS arm64.
• Linux and Windows are not supported by 3.0.0-rc.3 until their isolated acceptance matrices pass.
• No third-party npm packages or Bash runtime are required.

Future OpenClaw releases may work only when the public CLI fields still match the validated contract. Unknown fields fail closed.

Durable model

• memory/YYYY-MM-DD*.md — immutable daily source logs; never move, delete, or edit.
• memory/<topic>.md — durable L2 topic detail.
• MEMORY.md — compact current-state index and L2 pointers.
• memory/dream-log.md — human-readable diary, at most 30 entries.
• logs/signal-dreaming/state.json — rebuildable hash cursor outside memory indexing.
• .backup/memory-dreams/<run-id>/manifest.json — run status, plan, hashes, and backup map.

Do not depend on recall JSON. Detect work from daily-log SHA-256 changes, including same-day appends and suffixed logs.

Workflow

Read references/dream-protocol.md completely before a write run.

1. Run preflight.mjs.

2. Generate and save a read-only plan:

   bash

   node <SKILL_DIR>/scripts/delta-state.mjs plan <WORKSPACE_ROOT> > <PLAN_FILE>
   

   The planner combines changed daily logs and index maintenance. Stop only when "noop": true. A hard-limit index must produce a non-noop compact-first plan even when no daily log changed.

3. Read selected daily logs and relevant L2 files. When plan.index.maintenanceRequired is true, prioritize semantic compaction: keep current state, recovery conditions, authority boundaries, and L2 pointers; move detail to L2 before removing it from the index.

4. Begin the guarded candidate transaction:

   bash

   node <SKILL_DIR>/scripts/run-guard.mjs create-run-id <WORKSPACE_ROOT>
   node <SKILL_DIR>/scripts/run-guard.mjs begin <WORKSPACE_ROOT> <RUN_ID> <PLAN_FILE> MEMORY.md memory/dream-log.md memory/<topic>.md
   node <SKILL_DIR>/scripts/run-guard.mjs verify-before-write <WORKSPACE_ROOT> <RUN_ID>
   

5. Edit only the manifest's candidate files under .backup/memory-dreams/<RUN_ID>/candidate/. Never edit live MEMORY.md, live L2, or the diary during the candidate phase.

6. Prepare the diary entry JSON and semantically review candidate L2 files.

7. Finalize:

   bash

   node <SKILL_DIR>/scripts/run-guard.mjs finalize <WORKSPACE_ROOT> <RUN_ID> <ENTRY_JSON> --semantic-review-confirmed
   

Finalization audits the candidate, rechecks every live hash, renders the diary in the candidate tree, then commits planned files and advances state. Candidate rejection leaves live memory untouched. Omit --semantic-review-confirmed only when no L2 candidate is touched.

Limits

• Missing state: process only the most recent 7 calendar days.
• Per run: at most 32 daily logs and 512 KiB total input.
• --full-history is manual only and remains bounded.
• MEMORY.md at 0–8192 bytes is healthy.
• 8193–10240 bytes is a soft trimming band.
• More than 10240 bytes is a hard failure for candidates and a mandatory compact-first trigger for the planner.
• A soft-band index triggers maintenance when it changed since the last successful review or that review is at least 7 days old. An unchanged recently reviewed soft-band index does not loop nightly.
• A maintenance-only candidate must reduce MEMORY.md; otherwise reject it without writing a diary or advancing state.

Failure and recovery

Manifest states distinguish candidate failure from uncertain live state.

• candidate_rejected: deterministic candidate failure; live memory was not changed, the lock is released, and no acknowledgement is required.
• incomplete: a live commit may be partial or an older live-edit run needs reconciliation. Inspect backups and live hashes before acknowledgement.
• committed: candidate audit, live hash recheck, planned-file commit, diary, and state all completed.

Never start over an active, stale, or unreviewed incomplete run. For an incomplete run:

bash

node <SKILL_DIR>/scripts/run-guard.mjs ack-incomplete <WORKSPACE_ROOT> <RUN_ID> --confirm <RUN_ID>

Acknowledgement records review; it does not restore files.

Inventory retained V2 and V3 transactions without mutating them:

bash

node <SKILL_DIR>/scripts/transaction-list.mjs <WORKSPACE_ROOT>

The inventory accepts legacy V2 workspace/entries manifests and V3 root/plannedFiles manifests. It recognizes the legacy backup lock directory and the V3 logs/signal-dreaming/run.lock file, fails closed when both lock formats are present, and rejects mixed or conflicting manifest identities. V3 results are diagnostic only and never emit automatic rollback commands.

For STATE_INVALID or STATE_SCHEMA_UNSUPPORTED, inspect and quarantine only the exact reviewed state hash:

bash

node <SKILL_DIR>/scripts/delta-state.mjs inspect-state <WORKSPACE_ROOT>
node <SKILL_DIR>/scripts/delta-state.mjs quarantine-state <WORKSPACE_ROOT> --confirm <STATE_SHA256>

Run standalone audit and isolated acceptance tests with:

bash

node <SKILL_DIR>/scripts/dream-audit.mjs <WORKSPACE_ROOT>
node <SKILL_DIR>/scripts/self-test.mjs

Scheduling

Provide cron only as a template; never create or modify it automatically. Schedule one isolated agent turn and include the workspace root. The job must run preflight and stop on no-op or any ambiguity.

json

{
  "name": "signal-dreaming-v3-daily",
  "schedule": { "kind": "cron", "expr": "<MINUTE> <HOUR> * * *", "tz": "<TIMEZONE>" },
  "sessionTarget": "isolated",
  "payload": {
    "kind": "agentTurn",
    "timeoutSeconds": 900,
    "message": "Run signal-dreaming as a scheduled invocation. Read its SKILL.md and references/dream-protocol.md completely. Workspace root: <WORKSPACE_ROOT>. Run preflight with --scheduled. Stop on preflight failure, no-op, ambiguity, or an unfinished run."
  },
  "delivery": { "mode": "announce", "channel": "<CHANNEL>", "to": "<TARGET>" }
}

For a v1.3.1 upgrade, leave existing memories and diary in place. Missing V3 state triggers the bounded 7-day bootstrap; diary numbering uses the maximum valid existing Dream number plus one.