An engineering session can end with a patch and still leave an important question open.
It may reveal a recurring decision, a missing contract, or a reliable validation step. The work is only compounded when that result reaches the surface that can carry it into the next execution.
The useful question is narrower:
Which existing surface can own, revise, and reuse this lesson?
The answer is a routing decision. It identifies who can revise the lesson, what can prove it still applies, and where the next actor will meet it.
A Capture Is Not Yet Memory
A transcript, command log, or session summary can preserve context. It can recover an unfinished question and explain why a decision looked reasonable at the time. It cannot, by itself, make the result reusable.
The durable form depends on what the lesson changes. A repository-specific contract belongs near the repository. A repeatable operator workflow may belong in a skill or runbook. A public research claim belongs in a note that names its evidence and its uncertainty. A mechanism belongs in the test, script, or workflow that can keep proving it works.
The destination matters because it supplies the revision path. A lesson that cannot be corrected by the people and systems closest to the work will eventually become stale advice.
Route By Ownership, Not Convenience
It is easy to put every useful observation in the most convenient place: one large context file, a personal note, or a growing collection of summaries. That makes capture feel complete while making later verification harder.
Route a lesson to the surface that can revise it:
- Repository behavior or contribution practice belongs in the repository documentation, tests, scripts, or workflows.
- A cross-repository operating contract belongs in a reviewed shared handbook or workflow surface.
- A repeatable personal or agent procedure belongs in a compact skill with source-backed references.
- A public claim that is still developing belongs in a research object, model, or dated research note.
- If nothing needs durable reuse, leave it as temporary execution residue.
These are accountability choices, not storage categories. The selected surface should be able to answer: who can revise this, what evidence supports it, and how will the next actor encounter it when it matters?
The Smallest Durable Layer
The smallest appropriate layer keeps the lesson close to the source that makes it checkable. It avoids copying private execution details into broad guidance. It also makes absence meaningful: a session may reveal no lesson that deserves promotion. That is a valid result.
This is why cleanup and capture need an ordinary sequence:
- Recover the candidate lesson from the trace.
- Check it against the source that owns the work.
- Put the reusable result in the smallest place that can maintain it.
- Keep or remove the trace under a stated review rule.
The trace can remain useful as history. The promoted form is what should carry the next execution. The two roles should not be confused.
A Test of the Engineering Memory Model
This is the first concrete test for the five-part engineering-memory model.
An old session supplies an observation. The current repository, workflow, or research source supplies evidence. Routing establishes the boundary: a lesson cannot become a general rule until an owner and revision path are clear. The promoted statement becomes the generalization. The selected documentation, mechanism, or research record becomes the application path for the next actor.
If any part is missing, capture should remain provisional.
A well-organized archive is not enough. A useful memory system must make it possible to find the right owner, check the source again, and improve the lesson without preserving every step that led there.
Preserve Patterns, Not Residue
The useful question is not whether every session should be retained forever. It is whether a verified pattern has a destination that makes the next execution safer, clearer, or less repetitive.
When it does, promote the smallest version of that pattern into the surface that owns it. When it does not, preserve the trace only for the period and purpose it still serves.
That is how engineering memory stays connected to operational truth: not by remembering everything, but by giving the lessons that matter somewhere responsible to live.