Memory Sleep - Agent Memory Consolidation

Biological brains consolidate memories during sleep. Agents don't sleep - but they should.

This skill implements a nightly consolidation cycle that transforms raw daily notes into clean, structured long-term memory. Result: each new session starts with compact, high-signal context instead of bloated raw logs.

The Problem

Agents accumulate daily notes (10KB+/day) that grow unwieldy. New sessions either:

• Skip reading old notes (losing context)
• Read everything (wasting tokens, hitting limits)
• Read partially (inconsistent recall)

This causes "session drift" - the agent forgets decisions, repeats mistakes, loses project context.

How It Works

Three-phase consolidation modeled on biological sleep:

Phase 1: NREM - Stabilize (extract what matters)

Scan daily notes from the last 3 days. For each entry, classify:

• KEEP - decisions, outcomes, new facts, action items, lessons learned
• DISCARD - routine logs, duplicate info, resolved issues, process narration

Phase 2: REM - Integrate (connect and structure)

• Update long-term memory file (MEMORY.md) with distilled insights
• Create or update per-project summaries if projects are tracked
• Remove outdated entries from long-term memory
• Resolve contradictions (newer info wins)

Phase 3: Cleanup

• Compress daily files older than 7 days into weekly summaries
• Remove entries that were fully integrated into long-term memory
• Generate session bootstrap summary (top priorities, open threads, key context)

Setup

Add a cron job for nightly consolidation. Use a cheaper model (Sonnet) to save costs:

Schedule: 0 4 * * * (4:00 AM daily)
Model: sonnet (or equivalent cheaper model)
Session: isolated
Timeout: 300s

Cron prompt template

Memory consolidation cycle. Read the skill file at [SKILL_PATH] and follow its phases.

Workspace: [WORKSPACE_PATH]
Daily notes pattern: memory/YYYY-MM-DD.md
Long-term memory: MEMORY.md

Phase 1 Instructions: NREM Stabilize

Read daily note files from the last 3 days (memory/YYYY-MM-DD.md).

For each section, apply the KEEP/DISCARD filter:

KEEP criteria (any one = keep):

• A decision was made ("we decided to...", "agreed that...")
• New factual information (credentials, URLs, contacts, project details)
• A lesson was learned ("never do X because...", "root cause was...")
• An action item exists and is not yet done
• Relationship/people context that aids future interactions
• Project status changes (launches, blockers, milestones)

DISCARD criteria (any one = discard):

• Routine operational logs (heartbeat OK, deploy succeeded)
• Process narration ("I then checked...", "running command...")
• Duplicate of information already in long-term memory
• Resolved issues with no lasting lesson
• Temporary debugging output

Extract KEEP items into a structured list grouped by project/topic.

Phase 2 Instructions: REM Integrate

Read the current MEMORY.md (if it does not exist, create it with a # MEMORY.md header before proceeding). For each extracted KEEP item:

1. Already in MEMORY.md? → Skip (avoid duplicates)
2. Updates existing entry? → Edit in place with new info
3. New information? → Add to appropriate section
4. Contradicts existing entry? → Replace with newer info, note the change

Per-project summaries

If the agent tracks multiple projects, maintain a compact summary per project:

• Current status (1 line)
• Key contacts
• Last significant event
• Next action / blocker

Staleness check

Remove entries from MEMORY.md that are:

• More than 30 days old with no recent references
• About completed/cancelled projects with no ongoing relevance
• Superseded by newer information

Phase 3 Instructions: Cleanup

Archive old daily files

For daily files older than 7 days:

1. Extract any un-integrated KEEP items (safety net)
2. Create/append to memory/weekly/YYYY-Www.md with a compressed summary
3. The original daily file can be kept or archived based on agent preference

Generate morning briefing

Write a memory/briefing.md file containing:

• Top 3 priorities (based on recency and importance)
• Open action items with deadlines
• Key context for likely conversations today
• Any warnings (approaching deadlines, unresolved blockers)

This file should be under 2KB - optimized for fast session bootstrap.

Metrics

Track consolidation effectiveness in memory/sleep-log.json:

{
  "lastRun": "ISO timestamp",
  "dailyFilesProcessed": 3,
  "itemsKept": 12,
  "itemsDiscarded": 45,
  "memoryMdSizeBefore": 8500,
  "memoryMdSizeAfter": 9200,
  "briefingGenerated": true
}

Research

• Sleep-like replay reduces catastrophic forgetting - Nature Communications, 2022
• Sleep prevents catastrophic forgetting in spiking neural networks - PLOS, 2022
• Can sleep protect memories from catastrophic forgetting? - eLife, 2020