Mindkeeper — Complete Agent Memory System#

Mindkeeper bootstraps a production-grade, 3-layer memory system for OpenClaw agents. It turns a fresh agent into one that remembers, learns, and organizes knowledge automatically.

Architecture Overview#

Three layers, each serving a distinct purpose:

text

Layer 1: MEMORY.md          — Tacit knowledge (preferences, lessons, patterns)
Layer 2: memory/*.md        — Daily notes (raw capture of what happened)
Layer 3: /life/ (PARA)      — Structured knowledge repo, QMD-indexed
         ├── projects/      — Active endeavors with goals and deadlines
         ├── areas/         — Ongoing responsibilities (no end date)
         ├── resources/     — Reference material, how-tos, patterns
         └── archive/       — Completed or inactive items

Supporting systems:

• QMD — Local semantic search across all layers (BM25 + vectors + reranking)
• Daily consolidation — Cron job extracts knowledge from sessions automatically
• Transcript verification — Echo and store voice transcripts for accuracy
• Sign-off routine — End-of-day consolidation, memory update, and shutdown

Setup#

1. Run the bootstrap script#

The setup script creates the directory structure and starter files. It never overwrites existing files.

Windows (PowerShell):

powershell

pwsh -File "<skill_dir>/scripts/setup.ps1" -Workspace "<workspace_path>"

Linux/Mac (Bash):

bash

bash "<skill_dir>/scripts/setup.sh" "<workspace_path>"

Where <workspace_path> is the OpenClaw workspace root (typically ~/.openclaw/workspace).

2. Configure QMD (recommended)#

QMD provides semantic search across all your knowledge files and session transcripts. See references/qmd-setup.md for the complete setup guide including Windows workarounds.

3. Set up daily consolidation#

A cron job reviews all sessions and extracts knowledge into your PARA structure. See references/consolidation.md for the exact cron command, prompt, and configuration options.

4. Enable transcript verification (optional)#

For voice-heavy workflows, add inline transcript echo and local storage. See references/transcripts.md for the two-layer approach.

5. Configure sign-off routine (optional)#

End-of-day routine: consolidation → memory update → status report → optional gateway stop. See references/sign-off.md for the complete workflow.

Working with the Memory System#

Writing to memory#

• Quick capture → Append to today's daily note ( memory/YYYY-MM-DD.md)
• Lesson learned → Update MEMORY.md with the distilled insight
• Project update → Update the relevant file in life/projects/
• New reference → Create a file in life/resources/
• Decision record → Use the ADR template in life/resources/decisions/

Searching memory#

Use memory_search to find information across all indexed files. QMD searches MEMORY.md, daily notes, and the entire /life/ tree.

Maintaining memory#

• During heartbeats: review recent daily notes, promote insights to MEMORY.md
• During consolidation: cron job handles session → knowledge extraction
• Periodically: move completed projects to life/archive/, prune stale content

Templates#

Starter templates are in the templates/ directory:

TemplatePurposetemplates/project.mdNew project file for life/projects/``templates/area.mdNew area of responsibility for life/areas/``templates/resource.mdReference material for life/resources/``templates/decision.mdLightweight ADR for life/resources/decisions/``templates/daily-note.mdDaily note for memory/``templates/memory.mdMEMORY.md starter with recommended sections

Copy and customize as needed. The setup script uses these templates for initial file creation.

Reference Docs#

DocumentRead when...references/qmd-setup.mdSetting up QMD search backendreferences/consolidation.mdConfiguring daily consolidation cronreferences/transcripts.mdSetting up voice transcript verificationreferences/sign-off.mdConfiguring end-of-day routine

Tips#

• MEMORY.md is your soul — Keep it curated. It's the first thing loaded each session.
• Daily notes are disposable — Raw capture, not polished writing. Speed > quality.
• PARA is flexible — Don't overthink categories. Move things as understanding evolves.
• Templates are starting points — Customize them to match your human's workflow.
• QMD fallback is automatic — If QMD goes down, builtin search takes over seamlessly.