AI Pulse · Under the Radar

Most AI agents remember things by turning documents and chats into embeddings, then retrieving snippets that look relevant. Lemmalog takes a different route. It uses Datalog, a rules-based query system, to store facts, track where those facts came from, and account for when they were true.

That matters because business memory is rarely just a pile of similar text. You may need an agent to know that a customer changed suppliers in March, that the information came from a signed contract, and that an older agreement is no longer valid. A rules-based memory system can reason over those relationships more directly than a basic retrieval setup.

The project is only three days old, so treat it as an early signal, not settled infrastructure. Still, the author has published benchmark results rather than just describing the idea. Lemmalog reports 0.463 F1 on LongMemEval while using 38 times less context than full-context baselines, plus 0.533 F1 on LoCoMo. It is also exposed as an MCP server, meaning AI tools that support MCP can potentially connect to it as a memory layer.

For operators, the useful question is not whether to replace your vector database next week. It is whether your AI work needs auditable, time-aware answers instead of plausible summaries pulled from similar documents. This is the kind of thing we build into an AI command centre, where agents need to work from governed facts, not just search results.