I research AI agent memory.
AI agents forget. Every session starts from zero, the decision the team made yesterday does not exist today, and the work stops when the session ends. I build the memory that keeps the work going: the same context for reading and for writing, whoever does it, a person or an agent. The target is a team running more than one agent, whose members keep losing context to each other.
A product decision should have gone into search. One document in four did not show up. The reason only appeared because I measured, and it was in two product decisions, not in the measurement. That is the kind of thing I write about here.
Research comes before claims. I publish the path to the number, including when the bug was in the instrument. When a number does not hold up, it comes down.
Author
Posts signed by him, newest first.
ResearchNovember 3, 20266 min read
To know whether memory improves, we needed a number that does not depend on opinion. We used each document as its own question. The result showed three bugs in the instrument before it showed a problem in the product.
Read postEngineeringJuly 5, 20266 min read
The architecture behind our retrieval pipeline: BM25 + dense vectors + RRF + cross-encoder. No LLM in the ranking path.
Read postProductMay 28, 20264 min read
Each agent starts from zero. Decisions made in one session vanish in the next. Here's how shared memory fixes that.
Read postTechnicalMay 25, 20263 min read
Model Context Protocol is how AI agents talk to external tools. Here's the developer-facing version.
Read postEngineeringMay 20, 20265 min read
After 6 months of dogfooding ValorBrain with our own team, here's what surprised us.
Read post