Engineering work rarely starts from zero, but software agents often do. A new session can see the repository in front of it while missing the decisions, tradeoffs, and operational lessons that shaped it. The result is familiar: useful discoveries are repeated, old constraints are rediscovered, and context that should compound instead disappears between tasks.
The reset problem
A commit records what changed. It does not always explain why the change was necessary, what nearby approach failed, or which boundary must remain intact. Pull requests and work items add more of that story, but the evidence is spread across systems and written for different audiences.
That fragmentation matters when an agent joins the work. Without durable context, it can read the current code yet still miss a known coupling, an environment-specific convention, or a decision that was already settled.
What durable context changes
EDRAY Tech is exploring a governed way to turn engineering history into useful working context. The idea is straightforward: gather evidence from the places where engineering decisions already live, organize it around the repository and task, and make the relevant parts available before implementation begins.
The goal is not to replace the source material. Commits, code, reviews, and work items remain the evidence. Durable context is an index and orientation layer that helps an engineer or agent find the right evidence sooner.
This changes the shape of a task. Known constraints can be loaded at the start. Prior incidents can inform verification. Cross-repository relationships can be checked before a local change is treated as isolated. The team spends less time reconstructing the same map and more time testing the decision in front of it.
Human authority remains the boundary
More context does not grant more authority. An agent may understand a system better and still lack permission to change a ticket, merge code, deploy a service, or expose private information. Those boundaries stay explicit.
The same is true for confidence. Historical context can guide investigation, but current code and current runtime evidence still decide whether a claim is true today. When evidence is missing, the correct result is uncertainty—not a plausible invention.
That combination is the useful one: durable context, traceable evidence, clear human direction, and verification proportional to the change.