Building with context For engineers
A good summary can lose the most important detail
A summary can preserve the topic while losing the condition that changes the next action. Check whether scope, current status, and verification survive—not only whether the output sounds faithful.
Small qualifiers carry operational meaning.
“Use this for now” leaves a future decision open. “The owners still need to confirm it” says the work cannot yet be treated as settled. A summary that retains the surrounding technical idea but drops those qualifications can look accurate until someone uses it.
This matters when one task becomes context for another. The next engineer may never see the original conversation. If the retained record removes the pending state, the agent can plan around a prerequisite as though it were already satisfied. There need not be an invented sentence for the result to become misleading.
What our retention experiment tested.
On 15 October 2026, we completed a controlled context-retention experiment with 50 cases in ten scenario families: 30 development cases, ten validation cases, and ten held-out cases. Five variants came from one source-backed episode; the other 45 were authored synthetic fixtures. They were correlated cases, not 50 independent customer histories.
The experiment exercised coding capture, preparation, sanitization, extraction, an in-memory context representation, and serialization. It did not test database persistence, live retrieval ranking, or native meeting and document connectors. A reviewed preparation candidate was retained, and a later candidate was stopped by a priority regression.
That boundary matters. The experiment tells us about the context delivered at the measured boundary. It cannot establish the quality of the whole product or a customer’s next implementation. Later persisted-path follow-ups were separate experiments with different preparation and must not be pooled into this comparison.
The average did not settle the decision.
The later candidate improved the combined development and validation aggregate. But its validation rating for preserving verification and outcomes declined from 88.6 to 79.4 on a model-rated 0–100 dimension. A fresh paired repeat showed the same direction, from 88.8 to 80.6.
We had defined a stopping rule before seeing the result: a confirmed decline beyond the allowed margin in a priority dimension blocks the candidate. That rule prevented a higher aggregate from making the decision for us.
The model scores were a signal to inspect, not proof of missing information. A source audit confirmed two concrete omissions: a record retained a requirement for owner confirmation but lost the report that confirmation was still pending; another retained a request to check consumers but lost the report that they remained unchecked. Other suspected constraint losses were not substantiated in the audited cases and remained judge disagreements.
A prerequisite is not the current status.
The distinction is easy to miss because the two sentences concern the same action. “Release after confirmation” defines a dependency. It does not say whether confirmation has happened. A useful retained record may need both.
The following is a simplified illustration of the failure mode, not a verbatim source excerpt.
| Retained statement | Missing information | Effect on the next task |
|---|---|---|
| Proceed after the owner confirms the mapping. | Confirmation is still pending. | The next plan may assume the dependency is already satisfied. |
| Check affected consumers. | The consumer check has not been performed. | A later summary may imply review coverage that does not exist. |
| Use the current approach. | The original choice was provisional. | A temporary accommodation can become a permanent rule. |
Audit a summary before reusing it.
Choose a record that will affect a real next action. Compare it with the relevant source, then ask whether these details survive with the correct meaning. Do not reward a detail that exists only in an upstream note if the next task never receives it.
- Applicability: which component, caller, version, or situation does the statement cover?
- Status: is the idea proposed, accepted, superseded, or still provisional?
- Completion: what has actually happened, and what remains pending?
- Verification: which result was observed, which was reported, and which is only requested?
- Support: can a reader inspect enough evidence to challenge the interpretation?
The next use is the reason to preserve it.
Our retained candidate still had weaknesses, including lost provisional wording and incomplete dates or source support. Selecting it did not mean retention was solved. The lesson was narrower: a better-looking summary must not erase a condition that changes what somebody should do next.
For your next handoff, pick one pending item and one temporary decision. Ask another reader to identify them from the retained context alone. If the reader cannot tell what remains unresolved, fix the record before treating it as reusable understanding. Supporting source packets for our experiment remain private; the methods and limitations above describe the scope of the reported result.
Evidence notes
Prepared with AI assistance and reviewed for source fidelity and practical usefulness. Examples marked fictional are illustrations, not customer results.