01High risk
Stale embeddings
Vector entries continue representing source material after it changes or disappears.
Cleanup: Compare embeddings with source revisions, then refresh or delete mismatches.
Field guide
02High risk
Old prompt versions
Deprecated instructions remain connected to live workflows or copied into new ones.
Cleanup: Version prompts, map active consumers, and retire superseded copies.
03Medium risk
Agent logs
Tool calls, transcripts, and execution traces accumulate without a defined review period.
Cleanup: Set retention windows based on debugging, security, and compliance needs.
04Medium risk
Model cache folders
Downloaded weights and inference caches stay on machines after projects move on.
Cleanup: Inventory active models and remove unreferenced versions from managed storage.
05Low risk
Duplicated generated files
Agents create repeated exports, drafts, screenshots, and transformed assets.
Cleanup: Keep authoritative outputs and delete reproducible or superseded copies.
06High risk
Outdated context windows
Saved history includes instructions or facts that no longer match the current task.
Cleanup: Summarize useful state, discard obsolete state, and restart context deliberately.
07Medium risk
Local AI artifacts
Indexes, checkpoints, traces, and scratch files remain scattered across workspaces.
Cleanup: Document artifact locations and add project-level cleanup commands.
08Medium risk
Orphaned Docker volumes
AI services disappear while their models, databases, and transient data remain.
Cleanup: Match volumes to active containers before pruning those without owners.
09Low risk
Package cache from AI builds
Repeated experiments leave large package and compiler caches on developer machines.
Cleanup: Measure cache size and clear safe, reproducible content on a schedule.
10Low risk
Forgotten test outputs
Browser traces, videos, fixtures, and model evaluations persist after their run.
Cleanup: Keep failure evidence selectively and expire routine passing artifacts.
Risk levels are directional prompts for review, not universal severity ratings. Context, access, and downstream dependencies determine the actual risk.