Summary of "OpenClaw + Obsidian gives you super powers"

Overview

The video presents a way to fix OpenClaw (and also Hermes) agent memory by using Obsidian as a persistent “vault,” implementing a 4-layer memory architecture.

Core idea: add an Obsidian-based memory layer

The creator says OpenClaw’s memory is unreliable, and their approach makes it “basically flawless.”

Layer 3 is added as an Obsidian vault that stores session knowledge in Markdown files.
Unlike some existing layers, the Obsidian vault is not fully injected into every prompt (to avoid context overload). Instead, it is loaded selectively.

Existing OpenClaw memory layers (already present)

Layer 1: Built-in memory
- Contains essential facts (e.g., agent name, important local file locations).
- Is automatically included in every prompt.
agents.md / soul.md
- agents.md: rules for the agent.
- soul.md: “personality” / communication style instructions.
Session search (Layer 4 mentioned as built-in)
- Keeps a record of past sessions / cron jobs.
- The creator warns that if things get slow, it may be because Layer 4 gets overloaded.

New Layer 3: Obsidian Vault (key features)

The agent uses the vault in two main ways:

Session start “memory check” (loaded at the beginning of a session)
- At the start of each session, the agent searches the vault for what it needs about that session.
- Memory is pulled exactly when needed, avoiding constant context bloat.
Memory on demand
- If you ask something like: “Work on the project from 3 days ago”, it looks up the content from the relevant daily logs in the vault and resumes.

Specific problem addressed: compaction forgetting

The creator claims a prior issue: during compaction, OpenClaw may “forget things from 5 seconds ago.”

With this system:

When compaction happens, the agent checks the Obsidian vault for what was discussed right before the compaction, so the user doesn’t notice gaps.
They report seeing no compaction problems since using the setup.

Additional vault-backed structures

The system also introduces other vault-backed elements:

Mistakes file
- When the user calls out errors (example shown: “Hey OpenClaw… you did this wrong”), the agent logs the mistake.
- Intended to help improvement (“it knows what it did wrong”).
Working context files
- Supposed to provide a dynamic context file based on what the agent needs.
Agent shared workspace (shared directory)
- A workspace accessible to all agents (e.g., Hermes + OpenClaw).
- Enables cross-agent continuity: if Hermes creates/saves a project (like scripting a YouTube video) into the shared folder, OpenClaw can later find it and continue.

Setup guide / tutorial prompt (implementation steps)

The creator provides a copy/paste prompt to set up the full system:

Install Obsidian
Use the provided prompt so the agent creates:
- the 4-layer memory instructions
- the rules file (including instructions for storing/using memories)
- the shared workspace configuration

Important setup caution (validation step)

After running the setup prompt, they recommend verifying that:

memories are actually being stored in the vault
the agent wrote the needed instructions into agents.mmd

Some model variants may fail to write everything immediately, so you may need to instruct the agent to “burn into your agents.mmd” how the system should work.

Extra tool: turning the vault into a personal wiki

Once the memory system exists, the creator suggests building a visual interface (a “personal wiki”), such as:

an “Alex Finn wiki” where memories (concepts, sources, people, etc.) can be browsed
avoiding manual searching through Obsidian

Recommendations / workflow advice

Use at least two agents (OpenClaw + Hermes) side-by-side for best results.
The shared workspace is the mechanism enabling their collaboration.

Main speakers / sources

Primary speaker: the YouTube creator (speaks as “I” and “my system”), building and demonstrating the OpenClaw + Obsidian memory workflow.
Referenced systems/tools: OpenClaw, Hermes, and Obsidian.