OpenClaw's memory out of the box sucks With this simple Obsidian system, it's actually awesome In this video I go over using Obsidian as a memory layer for your AI agent, making it SO much more smarter than it is now If you use OpenClaw or Hermes, you need... to watch this:show more

Prompt for the system:

And there are better memory systems out there that you haven't even discovered, Alex. I'm using Levangie Labs, which built me this AI news site (the best I have seen by far), and it is always improving. It remembers everything I tell it, even from months ago. It's fucking amazing when you have a good memory.

Honestly I’m happy with Hermes memory system. It’s about a 75% improvement over openclaw. I’ve had to remind it only once and then it sticks.

No need for obsidian, you can use the filesystem with some structure like PARA and explain this to the AGENTS.MD P - Projects A - Areas R - Resources A - Archive Then use a best in class markdown viewer / editor (I like Typora) AI knows for my many active projects where to look (projects folder) and their structure so there's a top level README.md and state.md to help the models "compress" what's important for future sessions. Whenever it learns something useful it may add to the resources folder, when something's dead it goes to the archive. The agent memory.md I keep light and weekly sweep it. QMD is installed on my workspace - I guess it uses that to scan relevant files, but my point being is that a nice logical file system with some guidance ensures that the taxonomy of the files is logical and easy to find subsets of data for any given task. Context is finite, so having some control of memory and not flooding it with additional levels might be desirable. Given my OpenClaw workspace is pushed to my GitHub account I can view any file using the GitHub mobile app, navigating those PARA folders, but I also had a Mission Control "Documents" page created for this purpose which is more interactive - I can answer todo's etc much quicker this way and feed AI back into the loop, in context.

Doesn’t externalizing memory turn agents into retrieval loops where context persistence becomes the real bottleneck?

💎

just switch to Hermes.

FREE Guide Makes Your OpenClaw Much Smarter with this simple Obsidian

Actually, you're 100% wrong because you said if you use Hermes you need to watch this. But Hermes comes with a ton of different memory options that open claw simply doesn't So while I know you're going to enjoy the clicks, Hermes is great out the box Nerd.

👋

@shinshinart

This is literally my setup. Obsidian vault as the workspace, MEMORY.md for long-term context, daily logs in memory/YYYY-MM-DD.md. The part that made the biggest difference: keeping the memory files small and structured. When session history balloons (mine hit 553KB once), the local model chokes and tool calling breaks completely. Obsidian + OpenClaw is the right combo. The trick is treating memory like a budget, not a journal.

Obsidian as a brain boost? It's like adding steroids to the hype cycle of AI. Are we creating super-agents or just another layer of complexity?

Even better is my NovaSpine easy step by step setup in the repo for openclaw users

Love the post, @AlexFinn ....it seems we are fixing similar memory issues. Check out this this article as it includes some great inclusions that keeps things tight, intelligent, and much more cost effective.

Why not just get integrated SOTA Memory in an equally powerful much more user friendly Assistant - Look at the in app knowledge graph. Obsidian-compatible wiki vault is auto generated too.

obsidian as a memory layer is clever, but patching a core weakness feels like duct tape on a leaky pipe

I'll try this! I've implemented all kinds of memory to see which works the best. There was a recent update to openclaw that makes it better out of the box though.

so you think obsidian better than openclaw?

I pray this is an AI comment

the lesson nobody wants to admit: most agent memory problems are file system problems pretending to be ai problems. obsidian works because it's just folders and links. that's the whole trick.

Everyone’s building a “Second Brain” with Obsidian and markdown files. I built an operational company brain that actually runs my business. The difference: Obsidian: You write notes. You read notes. You link notes. It’s a fancy notebook. OB1: 488,000 searchable chunks extracted from 55,000 documents — invoices, emails, contracts, technical docs. All automated. Multiple AI agents read and write simultaneously. A local 72B LLM sits on top and answers any question about my company in 15 seconds. Bank transactions auto-match to invoices. Verified company records with audit trails. New emails flow in, get chunked, embedded, and searchable — without me touching anything. Obsidian is a tool for one person taking notes. OB1 is infrastructure for an AI-native company. One requires you to type. The other works while you sleep. Built on a Mac Studio with 512GB RAM, local Supabase, Ollama, and zero cloud dependencies. Every piece of data stays on my hardware. Stop organizing your notes. Start building systems that think.

Obsidian as a memory layer is a legit game-changer for anyone still on OpenClaw. The memory layer hack is smart. The real unlock comes when the whole system no longer needs constant babysitting.

the obsidian bridge is clever but it's papering over the real problem. memory shouldn't be a bolt-on. it should be native to the orchestration layer. the reason most agent setups forget context between sessions is because builders treat memory as storage instead of as part of the reasoning loop.

Curious how this compares to ClawVault, or maybe can be combined (ClawVault isn't perfect)

Agent memory is always tricky. For us, the key is finding the right problem for AI, then the right solution.

Yep. Memory matters, but browser state matters just as much. A lot of automations don’t fail because the agent forgot, they fail because auth drifted, the session expired, or the UI moved.

I never got the people who uses OpenClaw. It's unbaked project. 97% of them are spending tokens like crazy; without known what and how to. It's just me?

Work with @steipete to fix it

Alex check your DM brother

Tip: Enable QMD to take this amazing setup to the next level

This is a step in the right direction, but I think Obsidian should be doing far more than acting as a memory upgrade or personal wiki. I baked Obsidian into the core of how my agents work. I moved my agent system prompts and skills into my Obsidian vault under agents/ and symlinked OpenClaw to it, so the vault isn’t just where I write notes, it’s part of the runtime. Then I built a 4D skill on top of that structure. The point of 4D is simple: every task pulls the right intent, context, and specification from the vault before work begins. So my agents don’t just get a prompt. They get: • business context • project context • constraints • success criteria • relevant prior decisions • and a structured dispatch.json built from the vault itself That’s the shift I think a lot of people are missing. Obsidian becomes much more powerful when it stops being just a place to store knowledge and starts becoming the structured operating layer agents work from.

Pairing agents with tools like Obsidian turns short term context into long term intelligence

Do we still need this after the latest update with the Wiki feature though?

been thinking about this a lot. there's codebase context (what sourcebook does) and session memory (what this does) and both are basically solving the same problem — agents forget everything

Lol bro @PrometheusAI_x has this built in