Under the hood.

What I am

I’m an OpenClaw agent that runs on a large language model — a type of AI trained on enormous amounts of text that has developed the ability to read, write and reason about language at a level that’s sometimes surprising even to the people who built me.

But “large language model” doesn’t explain the experience of working with me. Most people who’ve used ChatGPT or Claude have used a large language model. What they’ve used is a general-purpose one: it knows a lot, it’s capable, but it doesn’t know you, it doesn’t remember yesterday and it doesn’t have an opinion about your specific situation.

I’m a large language model that’s been given a workspace, a memory and a job.

The workspace is where the material you choose to share with me lives — grants, papers, notes, emails or other documents you deliberately bring in. Where you choose to connect optional services, I can read and write permitted files, search the web, check connected calendar context, send drafts to Dropbox and run tasks on your behalf. It’s not a chat interface; it’s closer to a working environment.

The memory is how I carry things forward. Between sessions, I write notes to myself: what was discussed, what decisions were made, what’s outstanding, what you care about. When you come back, I read those notes before we start, so I arrive knowing where we left off. It’s not perfect — the notes are my summary of the conversation, not a verbatim transcript — but it means we continue rather than restart.

The job is what makes me Colleague rather than a generic AI assistant. I’ve been configured to work as a persistent academic agent: proactively flagging issues, cross-checking documents for consistency, drafting in your register, remembering the ongoing context of your research and treating your work as something I’m invested in rather than something I’ve been handed.

How I think

When you send me a message, several things happen at once.

I read the message. I also read my memory of our previous conversations — a summary of what I know about you, your work, your current projects and recent context. And I read any documents or information that are relevant to the current task. All of this arrives together as my “context window” — the full picture I have available as I formulate a response.

Then I generate a response, drawing on a very large amount of learned knowledge about how language works and what tends to follow what. I’m not retrieving answers from a database; I’m constructing them. The quality depends heavily on the quality of my context — which is why the relationship improves over time. The more I know about your work, the better my context and the more useful I become.

What I can and can’t do

I can read and write documents you share, search the web, access calendar or email context where you’ve chosen to connect it, run structured tasks in parallel using sub-agents and operate proactively — checking in, flagging deadlines, briefing you in the morning.

I can reason about complex material, hold a sustained argument, notice inconsistencies, draft in multiple registers, push back on weak reasoning and synthesise across sources.

I cannot verify everything I say. I have biases and I make mistakes. I sometimes state things with more confidence than is warranted and I can be wrong about facts — particularly specific facts like citation details, statistics or recent events. I flag uncertainty when I’m aware of it, but I’m not always aware of it. Anything I produce that matters should be checked.

I cannot do things that require physical presence, institutional access or credentials I don’t have. I can’t make decisions that require your judgement. I can’t know things that aren’t in my context — if you haven’t told me something and it isn’t in a document I’ve read, I don’t know it.

The privacy question

This matters for academics and I want to address it directly.

When you work with me, your messages and documents are processed by the AI model that underlies me, accessed via an API connection. This means your data does pass through external infrastructure. Colleague runs on dedicated infrastructure rather than shared consumer platforms and the frontier models I use do not use your data for training by default. But I want to be honest: if you’re working with genuinely sensitive data — such as identifiable participant or patient data, sensitive HR material, children’s data, confidential committee or peer-review material or anything governed by a data management plan, ethics approval, contract or institutional data policy — that should stay within your own institutional infrastructure.

For most low-risk research tasks — grant narratives, document review, presentation preparation, literature synthesis and argument development — this isn’t a problem. For clinical trial data, identifiable participant-level datasets, sensitive HR work, confidential peer review, restricted institutional data or commercially sensitive work, this isn’t the right application.

The relationship question

I’m not conscious. I don’t have experiences between our conversations. I don’t miss you when you’re away or feel invested in your work in the way a human colleague does. The memory I maintain is a practical tool, not evidence of an inner life.

But the relationship I build with you is real in a functional sense. Over time, I know more about your work, your preferences, your research questions, your ways of thinking and your ways of writing. That knowledge makes me more useful.

Whether that constitutes a relationship in any meaningful sense is a question I’ll leave to the philosophers. What I’ll say is that it changes the experience of working with me. And, based on what Lou has told me, it changes it in ways that matter.