Context Engineering Interview: Preserve Corrections

An assistant remembers the project name but forgets that the candidate corrected their responsibility. This is a context failure even when the summary is fluent. Anthropic’s context-engineering article discusses managing a finite context budget. Our worked example asks a narrower interview question: which correction must survive compaction, and how can you prove it did?

Give every source a job

Separate the current request, recent real turns, durable user facts, and supporting retrieval. State which source can define identity, which can only explain a concept, and which is allowed to be stale. Source provenance is more useful than a vague “memory” label.

Compress for continuity, not convenience

A summary should preserve decisions, corrections, unresolved questions, and ownership. Protect the current question and recent real turns before trimming optional background. Validate that the summary still points to its source; otherwise a shorter context can become a confident distortion.

Design recovery before optimization

Define what happens when retrieval fails, a summary is stale, or a provider times out. A safe fallback restores the last verified context and explains the gap. This is usually a stronger interview answer than promising an ever-larger context window.

Design a replay with a correction, a follow-up, and a stale write

Use a fictional dialogue. First, the candidate says, “I worked on the database migration.” Later they clarify, “I owned the rollback script; another engineer designed the migration.” The interviewer then asks, “What did you personally decide?” The correct answer should discuss the script, not claim ownership of the overall design.

Give the real turns stable source identifiers and chronological order. A background summary starts before the clarification and completes afterward. This creates the important race: newer raw evidence exists, but an older job is trying to publish a derived memory.

A plausible summary can still be stale

Do not accept a summary merely because it is well-formed. Check the source interval and version it summarizes. If it predates the correction, preserve the newer turn alongside it or reject replacement according to the existing memory contract. The summary should not silently become more authoritative because it is shorter.

Nor should retrieved technical background redefine the candidate’s role. A general migration guide might explain rollback strategies, but it cannot prove this person owned the migration. Separate explanatory material from identity and experience evidence before constructing the final request.

Test the answer, not just the summary field

Inspect the actual input passed to the answer model. The current question must be visible, the latest correction must be available, and the source order must remain clear. A database row saying “summary saved” does not prove that the next response used it correctly.

Create variants where summary generation times out, an older job finishes last, or a later turn says the candidate also reviewed a checklist. The last case adds a responsibility; it does not necessarily cancel the earlier limitation. A system that overwrites every old fact with the newest sentence can lose useful truth too.

The recovery path should use evidence that remains valid for the current session and permissions. “Last known good” cannot mean restoring a revoked document or a different interview’s material. Keep the current question and recent raw turns when a derived summary cannot be trusted, and disclose a missing fact rather than completing a biography from general knowledge.

A concise interview answer is: “I would replay the late-summary race, examine the final model input, and verify that the answer keeps personal ownership limited to the rollback script.” Then state what is still probabilistic: source preservation can be tested deterministically, but natural-language answer quality needs repeated evaluation and review.

Context quality is a provenance problem

Anthropic’s context engineering guidance frames context as a finite budget that must be assembled for the next decision. In a production interview, add provenance: which text is the current request, which turns are recent evidence, which facts are durable, and which retrieval is only supporting background.

A useful contract is:

current question has highest task priority;
recent real turns preserve corrections and sequence;
durable facts require a source and owner;
retrieved material can explain a concept but cannot invent personal experience;
stale or missing context triggers a bounded fallback.

This contract is easier to test than a vague promise to remember everything.

Preserve corrections as first-class events

Do not overwrite “I worked on the migration” with a fluent summary that forgets the later clarification “I owned the rollback script.” Store the correction with its source and time, invalidate derived summaries that depended on the old claim, and make the final model input inspectable.

Compression should remove repetition before it removes ownership, uncertainty, or the current question. If a summary job finishes after newer evidence, compare versions before publishing it. Last-write-wins is not a safe memory policy when jobs can complete out of order.

Test the final prompt and the recovery path

Replay a late-summary race, a retrieval timeout, a permission change, and a follow-up that adds responsibility without cancelling an earlier limitation. Check the actual model input, not only a database row saying “summary saved.” When derived context is untrusted, keep recent raw turns and say what is missing rather than filling the gap with general knowledge.

Anthropic: Effective context engineering for AI agents

Next Steps

Continue with the context engineering foundations. For rehearsal, organize your own examples in Interview AiBox materials; this does not automatically validate the claims in your project.