AI Pulse · Under the Radar

A quiet argument is forming around a practical question: where should an AI agent keep what it learns? One new project, lemmalog, has arrived as an MCP server built around Datalog, a rules-based way to store facts and draw conclusions from them. It includes stratified rules and provenance tracking, meaning a business can potentially see not just what the agent believes, but which facts and rules led it there.

That matters because agent memory is becoming a real operational problem. If an assistant handles customer cases, sales research, or internal reporting, you need it to remember the right things, update them safely, and explain why it acted on a piece of information. A vector store can retrieve similar text, but it is less suited to answering clear logical questions such as, “Which customers qualify for this policy, and why?” Datalog-based systems are aiming to make that kind of memory more inspectable and rule-driven.

But there is no settled answer. A competing view gaining attention argues that agent memory should simply be plain files, not graphs, databases, or elaborate memory layers. The argument is that teams are over-engineering a problem that can often be handled with readable notes, folders, and version control. The debate is active on Hacker News and still developing.

For operators, don’t rush to pick a side. Match the memory system to the job. Simple workflows may need files. Workflows involving policies, permissions, customer state, or auditability may justify structured facts and rules. This is the kind of thing we build into an AI command centre, where memory needs to be useful, visible, and governed rather than mysterious.