Hermes Agent and OpenClaw, two popular AI agent tools, are having a moment. Both are impressive, both are genuinely popular, and both promise some version of the same thing, which is the thing almost everyone actually wants from AI: better memory and context across your work, so the tool stops forgetting who you are and what you are doing between sessions.
This is a more technical piece than most of what we write, and that is on purpose. If you are the kind of operator who cares about how this actually works under the hood, read on. If you are not, the one-line version is this: before you buy a tool to fix your AI’s memory, look at your files.
What people actually want, and whether a harness delivers it
Strip away the excitement and the wish is modest and reasonable. People want an AI that remembers context across projects, so they are not re-explaining their world every morning. Agent harnesses — software that wraps an AI model so it can act on its own and remember things between sessions — pitch exactly this. Hermes Agent markets itself as the agent that grows with you, with a built-in learning loop that persists knowledge across sessions. OpenClaw runs as a persistent process that remembers between runs and acts on your behalf.
Will that solve your memory and context problem? Maybe — possibly even well. But at a high token cost (these systems re-read large amounts of your information constantly, and you pay for every read) and a level of operational complexity most small businesses cannot sustain. You are no longer using a tool, you are operating a system: an autonomous process with its own configuration, its own failure modes, and its own appetite for computing power, which you pay for. That is a real commitment, and for many businesses the juice is not worth the squeeze.
The part the pitch skips: the substrate
Here is the issue underneath the issue: the substrate — the underlying files and records your AI actually reads. An AI tool, agent or not, can only work with the information it can actually find and read. Its memory of your world is a reflection of how your world is stored.
For most businesses, how it is stored is the problem. Files live in three different clouds and a desktop. The current version of a document is whichever copy the last person happened to open. Decisions live in email threads and people’s heads. The same fact exists in four places, slightly different in each. When you point an AI at that, the drift you experience — the AI losing track of what you have already told it — is not the model failing to remember. It is the model accurately reflecting a scattered, contradictory source. Garbage organization in, drift out.
This is why an agent harness is so often the wrong first move. It does not fix the substrate, it inherits it. Pointing a sophisticated autonomous agent at a chaotic file system gets you a sophisticated autonomous agent that is confidently working from a mess. The memory and context problem you were trying to solve is still there, now wrapped in more cost and more complexity.
An AI consultant will tell you an agent is the answer and you no longer need traditional automations. The less glamorous question sits underneath: how does your organization actually structure the long-standing data AI touches? That question decides whether any AI performs — and it gets skipped, because fixing habits is harder than buying a tool.
What we built for ourselves
We will be specific, because specificity is the point. Our own knowledge base is not an agent. It is a disciplined substrate, built so that any AI tool we point at it gets a clean, consistent, current view of the business.
The tooling is simple: plain markdown files on disk (a text format any tool, and any AI, can read), version-controlled in git so every change is tracked and nothing is silently overwritten, synced through OneDrive so it is the same on every device. We edit in VS Code (a text editor) and read in Obsidian (a note-taking app) when a visual map helps or when adding notes on mobile, but those are just windows onto the same files.
The structure is deliberate:
- A single source of truth.
- An inbox for low-friction capture.
- A log of decisions we only ever add to, never rewrite, so the reasoning behind a choice is never lost.
- A standing weekly review that triages new notes and reconciles drift.
The rule that holds it together is simple. A piece of work is not finished until the knowledge base reflects it.
None of that is exotic, and that is the lesson. There is no model in it. It works with any AI tool, current or future. It is discipline, expressed as file structure. When we do bring AI agents and automations against it, they work well, because they are reading from something built to be read by any AI. The substrate came first, on purpose, and the agent is a later decision made against a stable foundation rather than a patch for an unstable one.
Why we wrote this
This is an edge piece, and we know who it is for. We do not expect a busy non-technical owner to read this and reorganize how all their files and information are stored by tomorrow. The lighter-weight version of this discipline (the tool inventory and the data classification reference, a single agreed place for things) is where most businesses should start.
But if this resonated, if you are the kind of operator who suspects your AI problems are really organization problems, you are exactly who we like to work with. The businesses that get durable value from AI are not usually the ones with the newest agent. They are the ones with their data in order down to the level of habit. Want help implementing a data management solution that works across AI tools? Book a free consultation.