An agent runs across multiple sequential workflow jobs and needs to preserve intermediate results (such as parsed dependency graphs) between jobs without re-computing them. The data is only relevant for the current feature branch and does not need to persist beyond a week. Which GitHub mechanism is best suited for this use case?

The GitHub Actions cache (actions/cache) is designed exactly for this use case: storing intermediate computation results that should be reused across jobs or workflow runs on the same branch. Keying the cache by branch name and a hash of the relevant input (e.g., the lockfile hash) ensures the cache is invalidated when the underlying data changes. The cache auto-expires after a configurable period (default 7 days of inactivity), which matches the stated retention requirement.

Option A (workflow artifact with 7-day retention) would also work but artifacts are better suited for final outputs and cross-run audit data, not intermediate computation caches. Artifacts do not support cache-style invalidation based on content hashes, and they accumulate on every run rather than being overwritten when fresh data is available.

Option C (commit to a cache branch) is an anti-pattern: it pollutes the repository history with machine-generated data, creates write contention when multiple workflow runs occur simultaneously, and makes the data permanent unless explicitly cleaned up.

Option D (repository variable via API) is limited in size (typically a few KB), not designed for arbitrary structured data, and updating it requires API calls with appropriate permissions that may not be configured for the workflow.

The Actions cache with a branch-and-content-hash cache key is the correct, purpose-built mechanism for inter-job short-term agent memory.