Memory is All We Need
For the last couple of months I've been working on a small open source project you may have heard of — Openclaw. It went uber viral everywhere so much so that Peter who created it was invited by OpenAI to lead the personal assistant efforts; Openclaw will become a foundation and continue to exist and be supported by OpenAI and others. Great outcome.
I built lobster, a workflow shell for openclaw during the early days and that's how I became a part of the org. Around the same time I had started contributing via small bug fixes and then features. That led me to thinking more and more about memory, recall, and how openclaw being open source and not reliant on any user data plays to its advantage like no other application does.
How does this thing work?
Before we talk about memory I'd like to walk you through how openclaw works, how its magical, and why it will always be free and open to use.
In simple terms, Openclaw is not a harness or a wrapper or an orchestrator. It is a tunnel that allows you to connect any mode of communication, to your model of choice. Everything else that powers it is built on top of the dozens of CLI tools that Peter wrote over the years. The application is simple - it uses pi.dev as the harness to talk to the model provider of your choice, allow you to switch to a different model provider without losing your session etc. All communication with the model providers happens via your own API key or OAuth token. The session data, user data, permissions, configurations, memory, and most importantly your agent's soul.md ALL live in the form of files in your machine. This means you can build your own openclaw if you don't want to use the repo we work on. That's the magic of it.
Veiled gardens not walled gardens
As a software builder, after the initial magic of clawdbot settled (ok, it never wears off), the thing that was the most fascinating was the numerous CLI tools that make it super powerful [You can use vanilla openclaw without any skills or plugins and it can still be more useful than claude or chatgpt because of its resourcefulness — browsing the web, repairing itself, figuring out what tools need to be installed and doing as much of it in a self sustained manner as possible].
Almost all of these APIs have existed forever. Even if the CLI tools like gogcli that make these APIs palatable existed prior to openclaw, the GUI experience of browsing gmail was just so much better that there really was no need for someone to wield the CLI. With openclaw, automation with human in the loop helps you actually move faster than the GUI. My friend spent time cleaning up old emails from years ago, for instance. Yes you could have written scripts to do this before your AI agent came into your life, but the agent makes it a lot easier thus making you a lot more willing to interact with these APIs through your agent.
I've become very open source pilled when it comes to AI agents. For years since the web came into existence, companies like Google, Microsoft, Apple and Facebook have enriched themselves beyond anyone's imagination with the help of user data. They use data about you to serve ads, sell clickstream data to other advertisers etc. Once the model providers entered the scene, they started doing the same. OpenAI has said they'll start serving ads. All model providers use the data you send to make their next iteration of models more powerful. Ads are great, I've bought very many things via ads. I have no problem with these companies making money off me.
However, things have changed this time. Data about me so far was not valuable to me. It was valuable to advertisers and for companies to build an algorithmic feed. But now, it is even more valuable to me as I can use it to teach my AI agents about me, my needs, and unearth interesting things via the vast amounts of data that have been and are being collected about myself.
This is why we need veiled gardens over walled gardens. There's fewer and fewer ways for me to access my data and metdata from companies like Google, Apple, Facebook, or Twitter. And even more so when it comes to data about me that Anthropic, OpenAI, or other model providers have because of the walled gardens they have built over time to protect their richest asset - user data and metadata.
The magic that openclaw unleashed was that if you have all the data about you on your machine (memory, sessions, user profiles), and a way to access data on demand from providers you use (CLI tools to access your google account, twitter, or facebook data), you can build extremely powerful personal assistants without developing a dependency on a single model provider or a company.
The day local models can be just as powerful as SOTA models of today isn't very far. In fact, there are openweight models that can be run on a beefy mac studio today. But it is only going to get more efficient and cheaper to run powerful models locally that can serve all your needs.
Even if you're not running local models having all the data about yourself, and a way to JIT access other data that you need — like emails and tweets, is super valuable. First, it's your data, the providers can benefit from the data because they collect and use it, but it belongs to you. Second, the ongoing evolution of an understanding of who you are by your personal assistant agents is extremely valuable.
What does a user friendly agentic future look like
The answer is deceptively simple: it looks like memory you own. Not memory trapped in a chat window that resets when you switch providers. Not memory held hostage behind an API that could change its terms tomorrow. Not memory scattered across a dozen walled gardens, each holding a shard of who you are while selling the reflection back to advertisers.
Memory that lives on your machine. Memory that accumulates across every conversation, every tool, every model you use — and follows you when you leave. Memory that makes your agent better not because a company trained on your data, but because you chose to teach it. We've spent two decades watching companies turn our data into their moats. The shift that Openclaw represents — and that the broader open source agent ecosystem is racing toward — is the realization that the moat was always ours to build. Your emails, your browsing patterns, your preferences, your writing voice, your decision-making history — that's not training data. That's identity. And for the first time, the tools exist to make it work for you instead of about you. The technical pieces — local models, portable session formats, standardized memory schemas — will come. They're engineering problems, and engineers are already solving them. The harder shift is cultural: convincing people that the data they've been giving away for free is now the most valuable thing they own.
Memory is all we need.