DeepSleep v3.0

Two-phase daily memory persistence for AI agents.

Like humans need sleep for memory consolidation, AI agents need DeepSleep to persist context across sessions.

What's New in v3.0

Change	Impact
🔴 Unified pack+dispatch cron	Single 23:50 job replaces two separate crons. Saves ~2-3万 token/day in cron startup overhead
🔴 Silent day fast path

Skips history pull on silent days. Saves ~5-8万 token/day when nothing happened | | 🔴 Smart晨报 delivery | Only sends morning briefs when there's real content, P0 due items, or stale OQs. Eliminates noise | | 🟡 Memory importance tiers | 🔴P0/🟡P1/🟢P2 per entry. Different retention windows (7/5/3 days). Phase 3 loads P0+P1 first | | 🟡 Cross-group correlation | Detects related topics across groups. "视频训练实验室 ↔ 千问微调实验室: both discussing Qwen-VL" | | 🟡 OQ health tracking | Tracks age of each Open Question. ⚠️ at 7d, escalation at 14d, archive at 30d | | 🟡 Schedule priority + reminders | P0/P1/P2 priorities with differentiated reminder frequency. P0 overdue = daily nag | | 🟢 dispatch_policy flag | Pack sets active/silent in daily file header; dispatch reads it to decide behavior | | 🟢 Memory quality audit | Weekly self-audit: compare raw conversation vs summary to catch missed info |

When to Activate

Activate when user mentions daily summary, memory persistence, sleep cycle, cross-session memory, morning brief, or nightly pack.

Architecture

CODEBLOCK0

Phases

Phase 1: Deep Sleep Pack (23:50)

1. 1. Lock date — Prevent midnight race condition
2. Auto-discover sessions — INLINECODE2
3. Silent day detection — If no real user messages, enter fast path (no history pull, <5000 tokens)
4. Silent Day Carry-Forward with Decay — 4-tier decay: Days 1-3 full, 4-7 slim, 8-14 minimal, 15+ archive to MEMORY.md
5. [Active day] Pull conversation history — Parallel batch INLINECODE3
6. Filter and summarize with importance tiers — 🔴P0 strategic, 🟡P1 important, 🟢P2 routine
7. Cross-group correlation — Detect related topics across groups
8. OQ health check — Track age, warn at 7d, escalate at 14d, archive at 30d
9. Schedule update — With P0/P1/P2 priorities, key-based dedup
10. Write daily file — Idempotent, with dispatch_policy flag and importance tiers
11. Self-check — Validate chat_id annotations, importance tags, OQ dates
12. Update MEMORY.md — Only for P0 items, with guard rails

Phase 2: Morning Dispatch (same session, immediately after pack)

1. 1. Dedup check — dispatch-lock.md prevents double-sending
2. Failure recovery — If pack somehow failed, do emergency mini-pack
3. Read daily file — Already in context from pack (skip re-read in unified mode)
4. Check dispatch_policy — active = send briefs; silent = update snapshots only
5. Smart send decision — Per-group: send only if new content, P0 due items, or stale OQs
6. Send per-group briefs — With importance-highlighted summaries, cross-group hints, OQ warnings
7. Schedule reminders — P0 due/overdue = 🔥 in brief; P1 due = mention; P2 = summary only
8. Write per-group snapshots — Tiered retention: 🔴7d, 🟡5d, 🟢3d. OQs and todos never expire
9. Write dispatch log — Including send/skip counts and policy
10. Write dispatch lock

Phase 3: Session Memory Restore (on demand)

When the agent receives a message in a group:

1. 1. Check memory/groups/<chat_id>.md — load if exists
2. Priority loading — If token budget is tight, load 🔴+🟡 entries first, skip 🟢
3. Check cross-group correlations — If this group has related groups, note it
4. Check memory/schedule.md — Any P0 items due today? Proactively mention them
5. If snapshot missing/stale (>48h) — Fall back to INLINECODE10

Phase 4: Memory Quality Audit (weekly, optional)

Once a week (suggest: Sunday heartbeat or dedicated cron), self-audit pack quality:

1. 1. Pick 2-3 random groups from the past week
2. Pull raw sessions_history for those groups
3. Compare raw conversation vs the daily summary that was generated
4. Check for:

- Missed important info — Decisions or lessons that didn't make it into the summary - Over-compression — Context that was lost in summarization - Misclassification — Items marked P2 that should have been P1/P0 - OQ staleness — Questions that resolved in conversation but weren't marked done

1. 5. Write audit findings to INLINECODE12
2. If patterns emerge (e.g., consistently missing technical decisions), update filtering criteria in pack-instructions.md

Daily Summary Template

CODEBLOCK1

Schedule File Format

File: INLINECODE13

CODEBLOCK2

Priority levels:

• - P0 — Must complete on due date. Reminded in morning brief + Phase 3. Overdue = daily nag
• P1 — Important, mentioned in morning brief on due date
• P2 — Low priority, noted in daily summary only

Per-Group Snapshot Format

File: INLINECODE14

CODEBLOCK3

Setup

1. Create unified cron job (v3.0: single job)

CODEBLOCK4

⚠️ Migration from v2.x: If you have two separate cron jobs (deepsleep-pack and deepsleep-dispatch), remove them and create the single unified job above:
CODEBLOCK5

2. Optional: Weekly audit cron

CODEBLOCK6

3. Enable cross-session visibility

CODEBLOCK7

4. Initialize schedule

Create memory/schedule.md with the table header.

5. Add Phase 3 restore to AGENTS.md

See Phase 3 section above for the standing order template.

Requirements

• - OpenClaw with tools.sessions.visibility set to INLINECODE17
• Single cron job using agentTurn mode (--session isolated) with INLINECODE20
• INLINECODE21 on cron job (dispatch sends messages directly via message tool)

Privacy Notes

• - Pack writes a daily summary that dispatch may broadcast to groups. Never include private MEMORY.md content.
• Each group only receives its own summary — no cross-group content leakage.
• Cross-group correlations mention only the topic connection, never specific conversation content from another group.
• MEMORY.md is updated separately and stays in the main session only.
• Per-group snapshots only contain that group's own conversation summaries.
• DM content: only "N DMs processed" in daily summary; details go to MEMORY.md only.

Known Gotchas

1. 1. Cron timezone: --tz Asia/Shanghai means cron expression is in LOCAL time. Don't convert to UTC.
2. Must use isolated agentTurn: systemEvent ignores timeoutSeconds (hardcoded ~120s). Use --session isolated.
3. Timeout must be generous: 900s (15 min). Real-world unified pack+dispatch with 8 sessions takes ~3-5 min.
4. Parallel tool calls: Batch sessions_history and snapshot writes in parallel.
5. Chat ID in headers: Every ### GroupName MUST have <!-- chat:oc_xxx -->. Missing = group loses memory.
6. Header must be exact: ## DeepSleep Daily Summary — no variations.
7. dispatchpolicy flag: Pack MUST set <!-- dispatch_policy: active|silent -->. Dispatch reads this.
8. Time perspective: Briefs use reader perspective: "yesterday did X / today do Y".
9. Midnight race: Lock PACKDATE at start, never re-fetch.
10. Dispatch dedup: dispatch-lock.md written AFTER sends complete.
11. Schedule Key dedup: Check existing keys before writing.
12. Silent Day Decay: 4-tier (1-3d full / 4-7d slim / 8-14d minimal / 15+d archive).
13. MEMORY.md guard rails: Append/update only. Never delete/restructure.
14. DM privacy: Daily summary is broadcast; MEMORY.md is diary. Don't mix.
15. Importance tiers: 🔴P0 retained 7d in snapshots, 🟡P1 5d, 🟢P2 3d.
16. OQ age tracking: All OQs must have [since: MM-DD, pending N days] annotation.
17. Cross-group correlations: Only mention topic keywords, never quote another group's conversations.

Failure Recovery

If pack fails or times out:

• - Unified mode: Dispatch won't run (same session). Next day's unified job will detect missing summary and do emergency mini-pack for the missed day.
• Manual recovery: Run dispatch-instructions.md independently — it has built-in pack-failure detection and emergency mini-pack.

Verification Checklist

CODEBLOCK8

Inspirations

Built with insights from the community: agent-sleep (multi-level sleep), memory-reflect (filtering criteria), jarvis-memory-architecture (cron inbox), memory-curator (open questions).