SPEAKER
Ran Aroussi
Founder, VarOps and Automaze. Creator of yfinance. Author of Production-Grade Agentic AI.
Ran Aroussi has spent 35+ years on both sides of the business and code line. He created yfinance, one of the most widely used open-source finance libraries in the world, with 50K+ GitHub stars and tens of millions of installs a month. He has written ad-tech infrastructure serving billions of impressions a day, and today runs VarOps and Automaze, a software delivery shop shipping client work with agents in production. He is the author of Production-Grade Agentic AI and Company-Scale Agentic AI, and hosts the Old School / New Tech podcast.
Upcoming sessions
See complete program4 November
11:05 AM · 30 min
Your agent isn't stupid, it's blind.
Distill, don't store: what we learned building memory for company-scale agents
Most agent memory is a pile of raw text with a similarity search on top. It works in a demo and falls apart the moment two sources disagree, a decision gets reversed, or one person isn't allowed to see what another person said.
We built the memory layer differently, and I'll go straight past memory 101 into the design decisions that actually moved the needle.
No raw storage. The retrieval unit is a distilled claim with provenance, not a chunk. Humans do not remember every word of conversations, but rather remember what established – so we built the system that way, and I'll show what that does to accuracy and to context budget.
Belief has a timeline. What did the organization believe last quarter, what superseded it, and when. Most systems just accumulate contradictions and let the retriever pick a winner at random. Detecting the contradiction is a feature, not an error.
Permission survives distillation, which is the hard part. Distillation destroys the source boundary that access control lives on. A claim assembled from four systems with four different ACLs has to answer differently for different people. This is the piece I've seen almost nobody address.
Agents write to it too. They're a producer, not just a consumer – and letting an agent write into shared knowledge poisons it fast unless the schema is human-authored and the instances are machine-populated.
I'll show the architecture, the benchmark methodology, our scores on LoCoMo, LongMemEval, and BEAM (1M/10M), and the design decisions I'd reverse. White paper published alongside the talk.
