fleet-memory-manager（车队内存管理器）

memory-manager

Upgrade your agent's memory from basic notes to a 3-layer production system with nightly consolidation. Based on patterns from running 7 AI bots in production for 30+ days.

Overview

Every AI agent has the same problem: it wakes up fresh every session with no memory of what happened before. The naive fix is a single MEMORY.md file — but that doesn't scale. After a few weeks, it's either a wall of text the agent ignores, or so curated it's missing operational context.

This skill installs a 3-layer memory architecture that mirrors how humans actually store knowledge:

Plus a nightly consolidation cron that reviews recent sessions and promotes important context up the layers automatically.

When to Use This Skill

Activate this skill when the user asks you to:

• Set up memory for their agent
• Upgrade from a basic MEMORY.md system
• Add nightly consolidation or memory cron
• Improve agent continuity between sessions
• Track active projects across agent restarts
• Set up the memory-manager skill

Setup Instructions

Step 1: Run the Setup Script

bash ~/.openclaw/skills/memory-manager/scripts/setup.sh

This creates the memory/ directory, copies templates into place, and prints next steps.

Step 2: Customize the Templates

After setup, edit these files in your agent's workspace:

1. MEMORY.md — Add your agent's existing long-term context
2. USER.md — Fill in who the human is, how they work, what frustrates them
3. AGENTS.md — Review the startup sequence (already wired for 3-layer loading)
4. HEARTBEAT.md — Configure which projects to monitor

Step 3: Configure the Startup Sequence

Your agent's AGENTS.md must include this memory loading sequence at the top of "Every Session":

## Every Session

Before doing anything else:
1. Read `SOUL.md` — identity and persona
2. Read `USER.md` — who you're helping and how they work  
3. Read `memory/YYYY-MM-DD.md` (today) — what happened today
4. Read `memory/YYYY-MM-DD.md` (yesterday) — recent context bridge
5. **If in MAIN SESSION** (direct chat with your human): Also read `MEMORY.md`

Why read yesterday too? Sessions started after midnight won't have today's file yet. Yesterday bridges the gap. This prevented context loss in our production fleet during late-night sessions.

Step 4: Set Up Nightly Consolidation

Add a cron job to consolidate memory each night at 2 AM:

0 2 * * * openclaw cron run memory-consolidation --model anthropic/claude-opus-4-5 --channel <your-main-channel-id>

The consolidation prompt to configure in OpenClaw cron:

You are performing nightly memory consolidation for this agent.

Tasks:
1. Read memory/YYYY-MM-DD.md files from the last 7 days
2. Read the current MEMORY.md
3. Identify: significant decisions, lessons learned, architecture changes, resolved issues, evolving patterns
4. Update MEMORY.md with distilled insights (add new, update stale, remove obsolete)
5. Check USER.md — update any new preferences or patterns you observed
6. Write a brief summary of what you consolidated to memory/consolidation-log.md

Be selective. MEMORY.md should stay under 500 lines. Quality over quantity.
Signal completion: reply "CONSOLIDATION_COMPLETE" to the channel when done.

Why 2 AM? Low activity period. The consolidation model has full context of the day's events. Running nightly (not weekly) means no single consolidation is overwhelming.

Memory Loading Rules

In Main Session (direct human chat)

Load all three layers:

SOUL.md → USER.md → memory/today.md → memory/yesterday.md → MEMORY.md

In Group Chats / Shared Channels

Load only layers 2 and 3 (no MEMORY.md):

SOUL.md → USER.md → memory/today.md → memory/yesterday.md

Why skip MEMORY.md in group chats? Long-term memory often contains personal context — private preferences, health info, financial details — that shouldn't leak into conversations with strangers. Daily operational notes are usually safe.

In Subagent / Worker Sessions

Load only today's notes:

memory/today.md (if relevant to task)

Subagents are ephemeral. Don't load full memory — it wastes tokens and context.

Daily Notes Format

Each memory/YYYY-MM-DD.md file follows this structure:

# YYYY-MM-DD

## Sessions

### [Time] — Session summary
- What happened
- Decisions made
- Tasks completed / in progress

## Active Projects

### Project Name
- **Status:** In Progress / Blocked / Complete
- **Last action:** What was done last
- **Next:** What needs to happen next
- **Blockers:** Anything blocking progress

## Context for Next Session

Key things future-me needs to know to pick up without re-explaining:
- [item 1]
- [item 2]

## Raw Log

(Less curated — dump things here that might matter)

The "Context for Next Session" section is the most important. Agents often end sessions mid-task. Without this section, the next session has to reconstruct state from scratch. Write it as if briefing a colleague who just joined the project.

Active Project Tracking

Projects move through these states in daily notes:

PLANNING → IN_PROGRESS → BLOCKED → REVIEW → COMPLETE

The heartbeat integration (see HEARTBEAT.md) checks active projects and surfaces blockers automatically.

MEMORY.md Curation Rules

Long-term memory should be:

• Curated, not comprehensive — the distilled essence, not raw logs
• Actionable — things that change future decisions
• Evergreen — not "what I did Tuesday" but "lesson learned from the Tuesday incident"
• Organized by topic — architecture, lessons, human preferences, recurring patterns

Things that belong in MEMORY.md:

• Architecture decisions and the reasoning behind them
• Lessons learned from failures
• Patterns in how the human works / thinks
• Important context about key projects
• Things you've been told to always/never do

Things that do NOT belong in MEMORY.md:

• One-off task completions
• Information that will be stale in a week
• Raw conversation transcripts
• Things already captured in USER.md

Heartbeat Integration

Add to your HEARTBEAT.md to enable project monitoring:

## Active Project Check

For each project in active_projects.json (if it exists):
1. Check last updated timestamp
2. If project hasn't been touched in >48h, surface it: "⚠️ [Project] hasn't been updated in X days"
3. If status is BLOCKED, surface blocker to human
4. If status is COMPLETE but not archived, prompt to archive

## Memory Health Check (weekly, Sundays)

1. Check MEMORY.md line count — if >500 lines, flag for pruning
2. Check memory/ folder — if daily files >30 days old exist, flag for archiving
3. Report: X daily files, MEMORY.md is X lines, last consolidation: [date]

File Reference