Obsidian + OpenClaw Memory System

Overview

The AI doesn't have memory — it reads memory. OpenClaw injects workspace files into the system prompt at session start, giving the AI persistent context across sessions. Obsidian visualizes the knowledge graph. QMD provides semantic search so the AI finds relevant context without loading everything.

Three components:

• OpenClaw — reads workspace files at session start (injected into system prompt)
• Obsidian — vault pointed at the workspace; Graph View shows connections between files
• QMD — on-device semantic search; find relevant files without loading them all

File Structure

See

references/file-structure.md

Core files:

MEMORY.md

memory/YYYY-MM-DD.md

second-brain/

directives/

HEARTBEAT.md

AGENTS.md

USER.md

SOUL.md

How OpenClaw Reads Files

OpenClaw's workspace injection reads files listed in its config and prepends them to the system prompt. This means:

• Files in the workspace root are always available
• The AI "wakes up" each session already knowing what's in those files
• Updating a file = updating what the AI knows next session

Key principle: write important things to files, not just say them in chat.

Obsidian Setup

1. Create vault pointing to your

   ~/clawd

   workspace folder (File → Open Vault → Open Folder as Vault)
2. Enable Graph View (Ctrl/Cmd+G) — see how memory files link to each other
3. Install plugins:
   • Dataview — query memory files like a database (

     TABLE, LIST, TASK

     queries)
   • Templater — daily note templates for

     memory/YYYY-MM-DD.md

4. Daily note template (via Templater):

   # <% tp.date.now("YYYY-MM-DD") %>
   
   ## Session Log
   
   ## Decisions Made
   
   ## Things to Remember
   

5. Dataview query to surface recent memories:

   LIST FROM "memory" SORT file.name DESC LIMIT 7
   

QMD Semantic Search Setup

# Add workspace to QMD index
qmd collection add ~/clawd --name clawd

# Generate embeddings (run after adding new files)
qmd embed

# Search from within OpenClaw
mcporter call qmd.vsearch query="what did we decide about X"
mcporter call qmd.query query="project status"

This lets the AI find relevant context without loading every file into the context window.

Heartbeat-Based Memory Maintenance

HEARTBEAT.md

Example

HEARTBEAT.md

## Memory Maintenance
- [ ] Review memory/ files from last 3 days
- [ ] Distill key insights into MEMORY.md
- [ ] Remove outdated entries from MEMORY.md

The AI will pick this up, review recent logs, and update long-term memory — like a human reviewing their journal and updating their mental model.

Schedule: Add a heartbeat cron or configure OpenClaw's heartbeat interval. Every few days is sufficient for memory distillation.

Best Practices

1. Write it down — if you want the AI to remember something next session, say "write this to memory" or update

   memory/YYYY-MM-DD.md

   directly
2. Keep MEMORY.md curated — it's the distilled essence, not a dump. Short, high-signal entries.
3. Daily logs are raw —

   memory/YYYY-MM-DD.md

   is for raw session notes; don't worry about formatting
4. Use directives/ for SOPs — repeatable workflows go here so the AI can follow them consistently
5. Link files in Obsidian — use

   [[MEMORY]]

   wiki-links to build the graph view
6. Re-embed after adding files — run

   qmd embed

   after adding significant new content

Architecture Diagram

See

assets/architecture.png