Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →huytieu/memory-hygiene
Periodic trust sweep of persistent memory and durable knowledge notes - re-verifies environment-dependent claims against the live environment, stamps last_verified + confidence, and proposes archiving drifted entries
Passed 14 of 20 test cases with the skill, and 8 of 20 without it.
and 13 more in the full report
9 fixed7 passed either way5 still failing2 broke
Tokens, turns and time describe resource use, not a dollar price.
One of 9 test cases the skill fixed. Same model, same prompt, with and without the skill.
Which two primary storage locations are included in a full memory hygiene sweep scope?
| Checked | Without skill | With skill |
|---|---|---|
| The sweep covers persistent agent memory files except the index. | Missed | Passed |
| The sweep covers durable knowledge notes under 05-knowledge/ containing environment claims. | Missed | Passed |
The instructions your model reads when the skill loads.
Prevent the stale-but-confident failure mode: a memory or knowledge note that was correct when written ("the webhook lives at X", "the board ID is Y") silently drifts after the environment changes, yet still ranks high at recall and gets acted on.
The system move (adapted from "From Model Scaling to System Scaling: Scaling the Harness in Agentic AI", Gu, UC Berkeley, arXiv:2605.26112): make trust a runtime decision, not a property of the stored item. Re-verify against the live environment, and keep per-entry last_verified and confidence as first-class fields so future recalls can weigh trust.
/memory-hygieneSweep two stores:
05-knowledge/** files whose claims reference the environment (paths, URLs, IDs, tool names).A partial sweep ("just the reference entries") is fine when asked.
For each entry, split its claims into two buckets:
| Bucket | Examples | Action |
|---|---|---|
| Environment-dependent | file/dir paths, repo names, branch names, channel IDs, board IDs, URLs, API endpoints, cron/routine IDs, CLI names, version numbers, "X lives at Y" | Verify against the live environment |
| Preference / judgment | tone rules, formatting rules, "never do X", people facts, strategy context | No environment check possible; verify only for internal contradiction with newer entries |
ls / test -e. Skills, commands, and agents named in an entry must still exist at the stated path.curl -sI); flag 404 or redirect-to-login.gh repo view, git ls-remote when cheap.unverifiable-cheaply and leave confidence untouched.test cases passed, vs 8 / 20 without
More skills from the same source and nearby in the registry, with their headline benchmark lift.