The Problem Is Not Prompting. It Is Context Decay.

Prompting is the visible part of working with an AI system. Context decay is the part that compounds.

A good prompt can improve the next answer. It can frame a task, impose constraints, and prevent an obvious mistake. But software work is not a sequence of independent answers. It is a chain of decisions whose value depends on surviving the end of a session.

When the session disappears, the code diff remains. The reasoning that made the diff safe often does not.

That is a different problem from prompt quality.

What actually decays

The lost thing is not only memory of facts. It is the shape of the judgment around them:

• why a boundary exists;
• which trade-off was accepted;
• which apparently convenient pattern is forbidden;
• what evidence changed the plan;
• which uncertainty was left open rather than solved.

A later agent can read the resulting code and produce a plausible explanation. Plausibility is not recovery. It may describe what the code appears to mean while inventing why it was built that way.

This is why a project can accumulate implementation while losing intent.

Context has different homes

Once context engineering is treated as information architecture, “put everything in the prompt” stops looking like a solution. Different knowledge belongs at different distances from the work.

Immediate task context carries the decision being made now. Repository conventions carry recurring local rules. Architecture notes and decision records carry boundaries and rationale. Specs turn intent into something that can be reviewed before implementation. Tests encode behavior in a form the system can execute.

These are not interchangeable containers. A test can prove a behavior without explaining the product reason for it. A decision record can explain a trade-off without proving that the current code still follows it. A prompt can focus attention without becoming durable project memory.

The design problem is to put each judgment somewhere it can be found, checked, and revised.

Retrieval is not truth

External retrieval and local search help with continuity. They reduce the cost of finding relevant material and prevent every session from starting from zero. They do not automatically establish that the material is current, authoritative, or still owned by the team.

That distinction matters as AI systems become better at assembling context. A larger context can make an answer more informed while also making its provenance harder to see. The agent may have access to more text but no clearer answer to: who decided this, when, and what would invalidate it?

Durable context therefore needs more than availability. It needs a status, an owner, and a way to encounter disagreement.

Intent is the difficult category

Technical debt lives in code. Comprehension can often be rebuilt by reading the code with help. Intent is different. The goal, constraint, or reason for a decision may never have been externalized at all.

An agent can infer a rationale for a debounce, a tenant check, or an unusual data shape. It cannot know whether that rationale is the original intent, a historical accident, or a constraint that has since expired. Treating an inference as recovered intent is a quiet form of cognitive surrender: the system sounds certain precisely where the evidence is weakest.

The remedy is not to document everything. It is to externalize the decisions whose absence would make future work guesswork, and to give those artifacts a visible relationship to code and tests.

A project should remember selectively

The goal is not a giant instruction file or a permanently expanded context window. That simply moves noise into a new place.

A useful project memory is selective. It records rules that recur, decisions that constrain future work, assumptions that would be expensive to rediscover, and evidence that tells an agent when to stop. It also lets obsolete knowledge be marked as obsolete rather than silently competing with the current system.

Prompting still matters. It is the steering wheel for this task. It is not the archive, the constitution, or the owner of the road.

The harder question is now architectural: which judgments deserve to become project memory, where should they live, and who is responsible for keeping them true?