MemSWE
Agent memory systems are usually compared on different models, tools and tasks. MemSWE fixes the runtime, the model, the tools and the metrics, and changes only the memory, so the evidence chain can be inspected. Written with Eduardo García at ITESO.
- 360Verified executions across 15 tasks
- 6Memory conditions, each isolated and reproducible
- 3.8 s → 23.0 sLatency, from full context to the slowest memory system
- 0%On the 5 stress tasks under hidden verification. Reported as is.
Task
15 software maintenance tasks
Memory
6 conditions, one changes at a time
Fixed runtime
Same model, tools and metrics
Verify
Visible and hidden checks
Evidence
An inspectable chain of results
The MemSWE design in one line. Only the memory condition changes between runs.
Want something like this built for your company?
Book a callMessage me on LinkedInSee Agentic Engineering