Fleet 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 productio...

Security Scan

The skill's files and setup script are coherent with a 3-layer memory system, but some runtime instructions (automatic loading, nightly consolidation posting to a channel, and guidance about not asking permission) create realistic privacy/exfiltration risks that users should understand before installing.

Purpose & Capability

The skill's name/description (layered memory + nightly consolidation) matches what the files and setup script do: copy templates, create memory/ daily files, and provide prompts for consolidation. No unrelated binaries, env vars, or external downloads are requested. It's coherent for the declared purpose.

Instruction Scope

Runtime instructions explicitly direct the agent to read local files (SOUL.md, USER.md, memory/*.md, MEMORY.md), run a nightly consolidation that reads multiple daily notes and writes updates to MEMORY.md, and post a 'CONSOLIDATION_COMPLETE' message to a channel. AGENTS.md contains the instruction 'Don't ask permission to load memory. Just do it.'—this grants the agent broad discretion to load potentially sensitive local context without prompting the human. The consolidation prompt directs the agent to produce updates and send a completion signal to a channel; if the channel is public or shared, that can leak sensitive consolidated content. The instructions also reference scanning external channels (Discord), checking email/calendar, and notifying channels, which require external credentials that the skill doesn't request (it assumes other integrations exist) — this is expected but important to note.

Install Mechanism

No install spec or remote downloads; risk is low. The only code is a local setup.sh that copies templates and creates files in the user's workspace and memory/heartbeat-state.json. The script does not contact remote endpoints or modify system configs beyond file creation. This is a low-risk install mechanism.

Credentials

The skill requests no environment variables or credentials, which is proportionate. However, many of the suggested runtime actions (posting to a channel, scanning Discord, reading emails/calendars, invoking an OpenClaw cron with a model and channel) imply the agent will use other integrations that require credentials; the skill leaves it to the environment. Users should ensure those integrations are scoped and private. Also templates and README encourage storing local 'tools' notes (API keys, SSH details) — storing secrets in workspace files would be risky.

Persistence & Privilege

always:false (normal). The skill writes files into the agent workspace (expected for a memory system). It does not modify other skills or system-wide settings. The combination of autonomous invocation (platform default) plus instructions to auto-load memory without asking increases blast radius for privacy-sensitive content — notable but not an inherent install privilege escalation.

Scan Findings in Context

[pre-scan-none] expected: No regex-based pre-scan findings were reported. That is consistent with this being primarily an instruction/template skill plus a benign setup script. Absence of findings does not eliminate the behavioral concerns noted above.

What to consider before installing

This skill appears to implement the described 3-layer memory system and will copy template files into your agent workspace and create daily note files. Before installing: 1) Inspect USER.md and MEMORY.md templates and avoid putting secrets (API keys, passwords, financial or health data) into them. 2) Choose the cron/channel target carefully — the nightly consolidation posts status and could send consolidated content into whatever channel you supply. Use a private channel or dedicated admin-only channel to avoid leaking private memory. 3) Remove or edit the AGENTS.md line 'Don't ask permission to load memory. Just do it.' if you want the agent to prompt before reading or sharing private files. 4) Ensure any integrations the agent will use (Discord, email, model API) have minimal scope and credentials stored securely (not in the memory files). 5) If you want stricter privacy, avoid enabling automatic nightly consolidation until you audit what the consolidation prompt will write and where it will post. If you want, I can produce a hardened version of the templates/cron prompt that avoids posting consolidated content to channels and requires explicit approval before updates.

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:

Layer	File	What goes here
1 — Long-term	MEMORY.md	Curated wisdom, architecture decisions, hard-won lessons
2 — Operational	memory/YYYY-MM-DD.md	What happened today, active project state, raw context
3 — Tacit	USER.md	How your human works, preferences, frustrations, patterns

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

File	Location	Purpose
SKILL.md	~/.openclaw/skills/memory-manager/	This file
setup.sh	scripts/setup.sh	One-command installer
templates/MEMORY.md	templates/	Long-term memory template
templates/AGENTS.md	templates/	Startup sequence template
templates/USER.md	templates/	User profile template
templates/HEARTBEAT.md	templates/	Heartbeat config template