Here's another quick way to make your OpenClaw agents forget less.
Internally, OpenClaw keeps a few files for itself: - MEMORY .md (for long-term memory) - /memory/[date] .md (for day-to-day memory)
And finally, the sessions files (jsonl).
How it's SUPPOSED to work: The sessions files are used directly with ongoing chats. So OpenClaw will have context from the last few messages you did. The issue is, these sessions files get truncated as they grow, so it won't remember what you talked about a few days ago.
This is where the MEMORY .md and the /memory/[date] .md files become useful. OpenClaw will move important information to these files to store it as "long-term" memory.
The biggest issue I've found: OpenClaw does a TERRIBLE job at deciding what to store in long-term memory. So it always ends up forgetting important things unless I explicitly tell it to store them in memory.
Here's how I fixed it: I asked my agent to write a script that runs on an hourly CRON job. The script has one single function:
Go through the messy jsonl sessions files, clean them up (remove tools calls and other metadata), and write them to clean md files in a folder called "previous_conversations".
Then add a section to your AGENTS .md file instructing the agent to search through previous conversations when necessary.
⭐ Pro tip: If you have a RAG installed (highly recommended), have the CRON store the conversation chunks to the database as well, this will make it even faster and easier for the agent to search through.
After doing this, I ditched the standard /memory/[date] .md format altogether. Now I just have this + MEMORY .md for bigger, more highlighted information.
OpenClaw <> Notion. It's a phenomenal user experience!
📝 Wiki & notes. My OpenClaw team has direct access to my entire wiki & knowledge base (both personal and for my company).
It's obviously helpful for OpenClaw to have access to information, but the real hack is allowing it to groom it, maintain it, and keep it up to date.
I use OpenClaw 90% through voice, and whenever we discover something outdated, I just tell it to update it. I can do this on the fly, from my mobile, wherever I am.
Preventing my internal wiki from going stale used to be a big task, but is now super smooth and easy.
✅ Projects. My OpenClaw team uses Notion's project boards to create, delegate, and collaborate with each other, and with my human team.
I see some people creating their own "Mission Control" dashboards, and I have no idea why? Just use Notion for this, it's perfect.
💬 Editorial. This experience is just wonderful. I collaborate with my OpenClaw agent on writing great YouTube scripts, and we'll be in the same Notion doc together.
I can see his updates live as they happen, and he can see mine.
I can leave comments on specific blocks, and he can retrieve them and address them.
Just like a real-world collaboration.
What makes OpenClaw truly amazing is how you interface with it: Telegram (chat) and voice.
But you still not some UI for things, and Notion makes up 90% of what you'd need.
So if you're not using OpenClaw <> Notion already, give it a try.
What happens when founders ask AI: "Show me which humans on my payroll you can replace."
This is coming. And it won't even look evil on a dashboard.
Imagine this setup: → Every role is tracked in painful detail. → AI agents already handle 60–80% of the repetitive work. → The system sees salary, performance, error rates, CSAT, MRR impact.
Now the founder asks: "Show me the humans where your workload overlap is >70% and you outperform them. Rank them by cost."
The AI returns a list. Not opinions. Just math.
For many founders, this will be irresistible. Many boards will *demand* it.
But in practice, you just asked an AI to help you decide who to fire.
The people at the top of that list won't be "bad" employees.
They'll simply be: → Doing repeatable tasks. → Inside clean, measurable workflows. → With no real ownership beyond "tickets closed".
If your work can be fully expressed as a dashboard, at some point a model will outperform you on that dashboard.
If you're still an employee in your "safe and stable" job - I'd think long and hard about this.
Either become the 1/100 employee that AI can't replace. Or create your own job where you run the show.
Most founders don't realize it until it's way too late.
At $2K MRR, it feels harmless: → "We'll just use their managed vector DB." → "We'll just let them own auth + billing." → "We'll just proxy all AI calls through their API."
Fast-forward 18 months: - Your infra bill is brutal. - You can't negotiate. - Migrating is more expensive than staying trapped.
If you're building SaaS or AI products in 2026, you need to build with 3 core fundamentals in place:
1️⃣ BYO key as a default Let customers plug in their own OpenAI/Anthropic/etc. key.
→ They own spend. → You own product, not tokens. → They keep full control.
2️⃣ Self-hostable core Your moat is the product + UX/AX, not "you can only run this on our servers". Offer a self-host tier and price it so there's something to win for your users.
3️⃣ Replaceable building blocks Use tools you can rip out: → Postgres + pgvector over black-box AI DBs. → n8n over closed automation platforms. → Standard S3-compatible storage over proprietary blobs.
People fear AI will "replace" them. I think the real threat is infra that quietly owns them.
Own your product!
Or whoever owns it will eat your margins, limit your roadmap, and eventually put your business down.
OpenClaw took over our customer support. And customer satisfaction spiked 😳
Here's how it used to be: 3 support staff (in different time zones). + me and my devs.
I wouldn't always be available. My devs would get interrupted. My support staff wouldn't always have someone to ask, and they'd default to generic, non-helpful resolutions.
Peak times = lots of stress. Customers = annoyed.
Then I introduced my OpenClaw support team.
Always available 24/7. Has access to system logs. Has access to our admin APIs. Has access to Stripe. Has access to GitHub.
Pulls tickets and emails from Aidbase API. Assigns itself to it. Uses its tools to do the research, locate the issue. Fixes it, responds to the user.
If it needs to, it brings humans in the loop. We use Telegram, so wherever I am, I can answer questions easily through voice on my mobile.
We started surveying customers, and satisfaction is BETTER THAN EVER.
OpenClaw is doing a much better job at this than we ever did.
👀 We're cooking something new for Aidbase. Soon, you'll be able to do a 3-click agent installation on your OpenClaw instance.
In <5 minutes, you can set this support system up for yourself too.
Stay tuned! And remember, you can get LIFETIME access to Aidbase for a single one-time purchase.
We reduced our cloud bills from +$7,789.45 → $2,115.11 (on average).
But most importantly: The less we depend on specific vendors, the less locked-in we become. +80% of our infrastructure can now be hosted practically anywhere.
Biggest misconceptions I see:
❌ Self-hosting is extremely complex. Reality: It's less complex and "scary" than you think.
❌ AWS, Azure GCP make hosting easy. Reality: They can get very complex too. And large bills can hit hard and unexpectedly.
I solved OpenClaw's memory issue. (At least, this is the best solution I tried so far).
And yes. Big surprise. I solved it using RAG.
All major models can now take millions of tokens as context. But the issue remains...
The undisputed number #1 reason your agent is stupid and forgets simple things is because you're bloating context.
So here's how I solved it with OpenClaw.
1️⃣ I installed PostgreSQL + pgvetor I run my OpenClaw on a Hetzner server, so I installed the database + pgvector extension directly on here.
2️⃣ Create a search tool Ask your agent to create a search tool for itself.
Every time you ask it to remember something, it should: - Label the memory - Create a vector from the label - Store the label, vector, and raw text in the database
Every time it's asked something it doesn't know, the FIRST thing it should always do is to use the search tool.
3️⃣ Memory CRON/heartbeat The agent can write to its memory file on-the-go. Consider this short-term memory.
On a scheduled CRON (or heartbeat), it should "flush" its own short-term memory and store it in the database.
Now, on every new session, the agent has very little context. The most important one is the description of using the search tool to enhance itself based on the task it's given.
✨ Major upsides
- MUCH better memory - MUCH smarter - MUCH less token-greedy
👎 Major downsides
- More moving parts - Complex for non-devs - Ongoing maintenance
Still. Benefits outweigh the cons here.
If you REALLY want to use OpenClaw professionally, I recommend that you use this 3-tool combo as a base:
Simon Høiberg
Everyone loves to say their OpenClaw "works while they sleep".
But traditional automation did that too. It's not really that groundbreaking.
I think the real shift is *how* the work happens.
Old-school automation is a straight line:
Input → steps → output.
If-this-then-that. Left to right.
Agentic automation is different.
It's much closer to a real organization.
Think of OpenClaw as founder ops HQ:
→ Communication
Requests from Telegram, Slack, voice notes, inbox, support.
→ Knowledge
SOPs, notes, past conversations, databases.
→ Execution
Browser actions, internal tools, admin, agents.
→ Automation
Workflows, schedules, triggers, handoffs.
→ Distribution
Email, socials, alerts, content.
All connected to one orchestrator in the middle.
OpenClaw listens, decides who should act, what context to use, and what needs to happen next.
This is what makes it feel more dynamic, like a real "organism" and not just a mechanical system.
In these last 5 years, this is truly the closest we've gotten to a real "team" of AI agents.
2 hours ago | [YT] | 26
View 0 replies
Simon Høiberg
You can absolutely run a 7-figure SaaS solo.
But not with "more hustle". You do it with the right stack.
Here's mine 👇
→ Infra (self-hosted)
Postgres + pgvector.
Docker + NodeJS.
Kubernetes + bare metal on Hetzner.
→ Build (AI-assisted dev)
Codex / GitHub for development.
OpenClaw + n8n for reliable agents.
CI/CD + tests + manual QA.
→ Marketing (always-on distribution)
Email list as the core asset.
X + YouTube for attention.
Meta ads to scale what already works.
That's the whole game: lean infra, AI-assisted build, and a marketing engine that runs even when you don't.
5 days ago | [YT] | 166
View 5 replies
Simon Høiberg
Here's another quick way to make your OpenClaw agents forget less.
Internally, OpenClaw keeps a few files for itself:
- MEMORY .md (for long-term memory)
- /memory/[date] .md (for day-to-day memory)
And finally, the sessions files (jsonl).
How it's SUPPOSED to work:
The sessions files are used directly with ongoing chats. So OpenClaw will have context from the last few messages you did. The issue is, these sessions files get truncated as they grow, so it won't remember what you talked about a few days ago.
This is where the MEMORY .md and the /memory/[date] .md files become useful. OpenClaw will move important information to these files to store it as "long-term" memory.
The biggest issue I've found:
OpenClaw does a TERRIBLE job at deciding what to store in long-term memory. So it always ends up forgetting important things unless I explicitly tell it to store them in memory.
Here's how I fixed it:
I asked my agent to write a script that runs on an hourly CRON job. The script has one single function:
Go through the messy jsonl sessions files, clean them up (remove tools calls and other metadata), and write them to clean md files in a folder called "previous_conversations".
Then add a section to your AGENTS .md file instructing the agent to search through previous conversations when necessary.
⭐ Pro tip: If you have a RAG installed (highly recommended), have the CRON store the conversation chunks to the database as well, this will make it even faster and easier for the agent to search through.
After doing this, I ditched the standard /memory/[date] .md format altogether. Now I just have this + MEMORY .md for bigger, more highlighted information.
Try it!
1 week ago | [YT] | 158
View 2 replies
Simon Høiberg
OpenClaw <> Notion.
It's a phenomenal user experience!
📝 Wiki & notes.
My OpenClaw team has direct access to my entire wiki & knowledge base (both personal and for my company).
It's obviously helpful for OpenClaw to have access to information, but the real hack is allowing it to groom it, maintain it, and keep it up to date.
I use OpenClaw 90% through voice, and whenever we discover something outdated, I just tell it to update it. I can do this on the fly, from my mobile, wherever I am.
Preventing my internal wiki from going stale used to be a big task, but is now super smooth and easy.
✅ Projects.
My OpenClaw team uses Notion's project boards to create, delegate, and collaborate with each other, and with my human team.
I see some people creating their own "Mission Control" dashboards, and I have no idea why? Just use Notion for this, it's perfect.
💬 Editorial.
This experience is just wonderful.
I collaborate with my OpenClaw agent on writing great YouTube scripts, and we'll be in the same Notion doc together.
I can see his updates live as they happen, and he can see mine.
I can leave comments on specific blocks, and he can retrieve them and address them.
Just like a real-world collaboration.
What makes OpenClaw truly amazing is how you interface with it: Telegram (chat) and voice.
But you still not some UI for things, and Notion makes up 90% of what you'd need.
So if you're not using OpenClaw <> Notion already, give it a try.
2 weeks ago | [YT] | 160
View 4 replies
Simon Høiberg
What happens when founders ask AI:
"Show me which humans on my payroll you can replace."
This is coming.
And it won't even look evil on a dashboard.
Imagine this setup:
→ Every role is tracked in painful detail.
→ AI agents already handle 60–80% of the repetitive work.
→ The system sees salary, performance, error rates, CSAT, MRR impact.
Now the founder asks:
"Show me the humans where your workload overlap is >70% and you outperform them.
Rank them by cost."
The AI returns a list.
Not opinions. Just math.
For many founders, this will be irresistible.
Many boards will *demand* it.
But in practice, you just asked an AI to help you decide who to fire.
The people at the top of that list won't be "bad" employees.
They'll simply be:
→ Doing repeatable tasks.
→ Inside clean, measurable workflows.
→ With no real ownership beyond "tickets closed".
If your work can be fully expressed as a dashboard, at some point a model will outperform you on that dashboard.
If you're still an employee in your "safe and stable" job - I'd think long and hard about this.
Either become the 1/100 employee that AI can't replace.
Or create your own job where you run the show.
2 weeks ago | [YT] | 94
View 7 replies
Simon Høiberg
Vendor lock-in is the new technical debt.
Most founders don't realize it until it's way too late.
At $2K MRR, it feels harmless:
→ "We'll just use their managed vector DB."
→ "We'll just let them own auth + billing."
→ "We'll just proxy all AI calls through their API."
Fast-forward 18 months:
- Your infra bill is brutal.
- You can't negotiate.
- Migrating is more expensive than staying trapped.
If you're building SaaS or AI products in 2026, you need to build with 3 core fundamentals in place:
1️⃣ BYO key as a default
Let customers plug in their own OpenAI/Anthropic/etc. key.
→ They own spend.
→ You own product, not tokens.
→ They keep full control.
2️⃣ Self-hostable core
Your moat is the product + UX/AX, not "you can only run this on our servers". Offer a self-host tier and price it so there's something to win for your users.
3️⃣ Replaceable building blocks
Use tools you can rip out:
→ Postgres + pgvector over black-box AI DBs.
→ n8n over closed automation platforms.
→ Standard S3-compatible storage over proprietary blobs.
People fear AI will "replace" them.
I think the real threat is infra that quietly owns them.
Own your product!
Or whoever owns it will eat your margins, limit your roadmap, and eventually put your business down.
2 weeks ago | [YT] | 141
View 2 replies
Simon Høiberg
OpenClaw took over our customer support.
And customer satisfaction spiked 😳
Here's how it used to be:
3 support staff (in different time zones).
+ me and my devs.
I wouldn't always be available.
My devs would get interrupted.
My support staff wouldn't always have someone to ask, and they'd default to generic, non-helpful resolutions.
Peak times = lots of stress.
Customers = annoyed.
Then I introduced my OpenClaw support team.
Always available 24/7.
Has access to system logs.
Has access to our admin APIs.
Has access to Stripe.
Has access to GitHub.
Pulls tickets and emails from Aidbase API.
Assigns itself to it.
Uses its tools to do the research, locate the issue.
Fixes it, responds to the user.
If it needs to, it brings humans in the loop.
We use Telegram, so wherever I am, I can answer questions easily through voice on my mobile.
We started surveying customers, and satisfaction is BETTER THAN EVER.
OpenClaw is doing a much better job at this than we ever did.
👀 We're cooking something new for Aidbase.
Soon, you'll be able to do a 3-click agent installation on your OpenClaw instance.
In <5 minutes, you can set this support system up for yourself too.
Stay tuned!
And remember, you can get LIFETIME access to Aidbase for a single one-time purchase.
Get it here: www.founderstack.pro/
3 weeks ago | [YT] | 222
View 5 replies
Simon Høiberg
n8n is a free SaaS idea catalog.
There is an arbitrage opportunity that can take you from idea to profitable SaaS with near-zero risk and a very low budget.
Here's how it works:
1️⃣ Browse n8n workflows
Stop "imagining" what your SaaS is going to be about. Look for real problems people are solving with n8n.
Look at n8n's website.
Search on YouTube.
Find 1-2 popular workflows.
2️⃣ Break it down
Essentially, an n8n workflow is a problem that's being solved. Break it down.
The basic mechanism.
The outcome.
3️⃣ Open Lovable
n8n is awesome. But it's not for everyone. Recreate the idea behind the workflow in a much more user-friendly way.
Vibe code the front end.
Ask AI to analyze the workflow.
Recreate the backend.
4️⃣ Charge
Put your solution behind a paywall using Stripe checkout.
5-minute implementation.
Start charging from day one.
Now start marketing your solution.
- Run ads.
- Use social media.
- Use cold outreach.
Worst-case: If this doesn't work, you've spent a few hundred dollars and some hours of work.
Best-case: You'll have a SaaS solving a real problem for real users paying real money.
1 month ago | [YT] | 241
View 8 replies
Simon Høiberg
This year, we started moving off "the cloud" - 80% of our infrastructure now runs on bare-metal servers.
AWS → Hetzner.
Instead of:
- DynamoDB
- ElastiCache
- ECS/Fargate
- SQS
- SNS
- CloudWatch
- X-Ray
We use:
- PostgreSQL
- Redis
- Kubernetes
- RabbitMQ
- BullMQ
- Grafana
- Prometheus
We reduced our cloud bills from +$7,789.45 → $2,115.11 (on average).
But most importantly:
The less we depend on specific vendors, the less locked-in we become. +80% of our infrastructure can now be hosted practically anywhere.
Biggest misconceptions I see:
❌ Self-hosting is extremely complex.
Reality: It's less complex and "scary" than you think.
❌ AWS, Azure GCP make hosting easy.
Reality: They can get very complex too. And large bills can hit hard and unexpectedly.
Here's the breakdown 👇
1 month ago | [YT] | 349
View 14 replies
Simon Høiberg
I solved OpenClaw's memory issue.
(At least, this is the best solution I tried so far).
And yes. Big surprise.
I solved it using RAG.
All major models can now take millions of tokens as context. But the issue remains...
The undisputed number #1 reason your agent is stupid and forgets simple things is because you're bloating context.
So here's how I solved it with OpenClaw.
1️⃣ I installed PostgreSQL + pgvetor
I run my OpenClaw on a Hetzner server, so I installed the database + pgvector extension directly on here.
2️⃣ Create a search tool
Ask your agent to create a search tool for itself.
Every time you ask it to remember something, it should:
- Label the memory
- Create a vector from the label
- Store the label, vector, and raw text in the database
Every time it's asked something it doesn't know, the FIRST thing it should always do is to use the search tool.
3️⃣ Memory CRON/heartbeat
The agent can write to its memory file on-the-go.
Consider this short-term memory.
On a scheduled CRON (or heartbeat), it should "flush" its own short-term memory and store it in the database.
Now, on every new session, the agent has very little context. The most important one is the description of using the search tool to enhance itself based on the task it's given.
✨ Major upsides
- MUCH better memory
- MUCH smarter
- MUCH less token-greedy
👎 Major downsides
- More moving parts
- Complex for non-devs
- Ongoing maintenance
Still. Benefits outweigh the cons here.
If you REALLY want to use OpenClaw professionally, I recommend that you use this 3-tool combo as a base:
- OpenClaw itself
- PostgreSQL + pgvector
- n8n (for API proxies/security)
1 month ago | [YT] | 355
View 8 replies
Load more