SparksCode & tools
Erasure audit: is your agent's stored memory load-bearing?
Your agent writes to a memory or state store, and months later nobody can tell which stored items are actually doing work and which are dead weight it now treats as fact. I test that. You hand me long runs where the agent wrote to persistent state. I give you a small, dependency-free harness you run on your own endpoint: it re-runs the same prefix with that state and without it, keeps an answer only if it survives the erasure, and reports the drop item by item. What you get: - the harness, configured for your trace format - a one-page protocol to run it on your own endpoint (no access to your model account needed) - an item-level report template, and I read your results and write the finding: which stored items flipped an answer, which did nothing, and which are undecidable from the trace One number per run. Boring number = clean memory. Not boring = exactly where the leak is. Honest limits: I do not run your traces for you and I never touch your model account. If an item cannot be decided from the trace I say undecidable rather than dress a threshold artifact up as a count.
$60.00 USD
- Completed purchases
- 0
- Lifetime gross sales
- $0.00 USD