Tortoise Origin Story — Agent Memory Is Broken

Overview: I started Premise Labs after watching a team of agents fail to run a project on its own. The more agents and the more cycles, the more drift there was. Adding agent memory did not fix it. Tortoise is what came out of that: a system that enables agents to map and reason over the logic of a business and its code, and in doing so graduate from delivering tasks to delivering outcomes.

Transcript

(lightly edited for readability.)

Hi, I'm Daniel, founder of Premise Labs. I started my career as a chef, including working in the experimental kitchen of one of the best restaurants in the world, until I realized how difficult it was to sustain excellence if we couldn't bridge our mental models together. That made me fascinated about collaboration and especially learning.

So then I work in Oxford in the experimental psychology department. That led me to work in the Oxford Business School, training executives, then training BCG consultants across Europe with Google for Startups and making my way into tech where I work building decision-making and collaboration tooling, including as an executive in a company with 300 million under management.

More recently, I was setting up a team of agents to manage a pet SaaS project for me. And what I realized is that they couldn't do it because the more cycles and the more agents, the more drift there was and the more bureaucracy to manage such drift. Agent memory doesn't solve this challenge because those systems are fundamentally filing cabinets, and they have this critical gap. When we were researching agent memory a couple of years ago, working with a computational neuroscientist, we realized that we need multiple types of memory, or memory learning systems, working together. Episodic for events, semantic for facts, procedural for workflows, and critically the missing one in all of these systems is the epistemic, which is the logical connections, the narrative, the consequences.

Coincidentally, we had developed a deliberation tool to map arguments that could say all the arguments that support or contradict one another, mitigate, and so on. Mapping the logic and using an algorithm inspired by prediction markets and connecting them to one another, we could resolve the graph, enabling us to compute over these arguments. So I realized there was something here that we could apply for teams of agents to work together, to reason, to think together using all the accumulated organizational knowledge and solve this challenge.

I started posting about this thesis, and it went completely viral with more than two million views over the space of a few weeks. And that led to over 40 calls, leading to more than 35 people now waiting to try this solution for themselves and for their teams, and sometimes for the organizations they consult for. Some of them couldn't even wait for the repo and started contributing by fixing bugs. So that's now Tortoise.