For the past year I have lived inside MindPalace, an open-source personal agent system that runs much of my working life: email, teaching, research experiments, a GPU cluster and a nightly report on what it did. It is retrieval-augmented generation without the vector database. Its memory is plain files, with a README in every folder serving as both index and summary, and the model retrieves by reading the index and choosing what to open, the way a person uses a library catalogue. This talk covers what a year of living in it taught me. On memory: why a hand-curated index can beat embedding search, and what keeping it in every request actually costs. On failure: real incidents, including bots that were running but not listening, agents whose waiting turned out to be spending, and a report that was too smooth to be true. On judgement: where the human must stay in the loop, and what that means for teaching students to build with agents. The protocols and methods I demonstrate are open source.
Anthony Bourached is Associate Professor of Machine Learning and Creative AI at University College London, in IOE Culture, Communication and Media and the UCL Centre for Artificial Intelligence. He co-founded Oxia Palus, whose AI reconstructions of lost and hidden artworks were exhibited at the Louvre and covered by the BBC, CNN, MIT Technology Review and Smithsonian Magazine. His research spans generative AI, art authentication and benchmarks for reasoning in AI systems, and he teaches agentic AI to creative students at UCL. For the past year he has run MindPalace, an open-source personal agent system that manages his research, teaching and correspondence. He holds a PhD in Machine Learning and Neuroscience from UCL.
