All writing

What we learned building Onetuza on our own memory layer

Shipping our first product forced real answers to problems that were theoretical until then. Notes from putting our own research under production load.

Mekhos AI Labs·August 2026·8 min read
What we learned building Onetuza on our own memory layer

What production exposes

Building a memory layer for a production system exposed problems that are easy to overlook in research environments.

Operational challenges

Persistence, retrieval latency, stale information, conflicting memories, and operational observability all become important once the system is handling real workloads.

Memory as infrastructure

These lessons have shaped how we think about memory as infrastructure rather than simply another feature attached to an agent.