SPEAKER
Veljko Kovac
Head of FDE · Cognee
Veljko is Head of Forward Deployed Engineering at cognee, where he leads technical implementation of agent memory for enterprise clients. He previously founded a venture-backed startup building a governance layer for AI agents, raising capital in London and San Francisco. Before that, he conducted graph neural network research with Prof. Pietro Liò at the University of Cambridge and Dr. Petar Veličković at Google DeepMind.
Upcoming sessions
See complete program4 November
1:55 PM · 30 min
Memory Without Deterministic Coordinates Is Just Hearsay
Coding agents now work on the same repositories across engineers and across weeks, and every session ends with a finding. Those findings are saved as sentences in private memory files, with nothing that says which function they concern or which version the agent was looking at. Two agents can investigate the same function a month apart, and no tool can tell their notes are related. This talk shows why the obvious fixes fail for the same reason, and why the one that works needs a deterministic code graph: stamping every finding, as it is written, with the identity of the function and the version the agent was looking at, computed the same way on every machine. Using a deterministic code graph as the coordinate system and a memory layer that stores findings against it, findings become shareable and comparable across agents and repositories, and summaries over them become counts rather than guesses.