What Termyte Can and Cannot Prove

Instrumentation is not causality

As founders, we have to separate a product being observable from a product being proven effective. Termyte can connect derived memories to observations and traces. It can record which context items were shown, preserve task evidence, and store feedback such as helpful, harmful, ignored, or corrected.

That gives you a trail to inspect. You can ask what was captured, what was derived, what was selected, and what feedback was recorded.

What the trail does not prove

A recorded context item is not proof that it caused a successful code change. A memory marked helpful is feedback, not a controlled experiment. A task that passed after context was delivered may have passed for several other reasons.

Retrieval ranking has deterministic bounds, but it is not calibrated on a public coding-agent corpus. We do not have a credible basis for claiming that one memory strategy universally improves agent performance.

The current product also does not provide comprehensive redaction, sandbox isolation, or a complete task-management interface in the viewer.

The research standard

Useful infrastructure should make its evidence visible without turning that evidence into a marketing claim. The research loop is straightforward: define a task, capture the available context, observe what was selected, record the outcome, and compare against a baseline.

Until that work is done on representative workloads, the honest claim is narrower: Termyte makes continuity inspectable. That is enough product surface to test. It is not enough evidence to claim causation.