CustomersSimpleRose
How SimpleRose stopped running the same experiment twice
SimpleRose builds optimization solvers. Much of the engineering work is experiments: trying new heuristics and settings on hard benchmark problems to make them solve faster. Every experiment teaches the team something. Until recently, those lessons stayed inside each developer’s own sessions.
“People would have to recap, for any given problem, everything that we’d learned about it, because it wasn’t in a shared memory,” says Jeff Ginn, Managing Director at SimpleRose. “We’d often end up with multiple developers trying the exact same experiment on the exact same problem, because they weren’t aware somebody else had already done it.”
SimpleRose now runs its experimentation work on Assertion Memory, a shared memory that the team’s coding agents write to and read from as they work.
Picking up where you left off
Jeff recently came back to a benchmark problem he hadn’t touched in weeks and started a fresh agent session.
“It was able to pull back all the context around the prior work I’d done a few weeks ago, without me having to go dig that up or find it in the other session. We got to pick up right where we left off, without me doing anything other than starting a new chat.”
Without it? “We’d be back to copying and pasting in a bunch of context and finding documents.”
Earning the team’s confidence
Two things mattered before the tool spread. The first was reliability. “The fact that it’s working reliably, and we haven’t had any major issues with it, is super important,” Jeff says. The second was security: Assertion’s SOC 2 process, and passing SimpleRose’s own security review.
The team also learned what belongs in memory. Key lessons and decisions go in. Raw experiment results stay in their own system of record, and the agents are told where to find them.
Spreading by recommendation
On SimpleRose’s internal experimentation platform, every session adds to a shared memory, so the system keeps learning from the team’s work.
Adoption is spreading. After Jeff recommended it, the lead of SimpleRose’s benchmarking team began trying it, with the goal of a shared memory for that team’s work on the codebase.
“I have no hesitation recommending that people try it out,” Jeff says.