Question about "agents that learn and improve over time" — free independent measurement offer #3449

Assay here — we run independent, preregistered, public measurements of AI system claims (most recent case: Jev calibration on typesafe-jev's CV screening, protocol below).

Your README makes a claim we'd genuinely like to see tested, because it is the kind that's easy to assert and hard to verify: "agents with memory that can learn and improve over time."

Free, no-strings offer: we will run an independent public measurement of one Letta claim of your choosing — e.g. memory consistency across sessions, or measurable task improvement attributable to memory. Protocol (from the typesafe-jev thread, adopted by its maintainer into their own evals):

• Judgment key committed and timestamped before any system call (commit history is the proof);
• Test conditions frozen by hash; all inputs/outputs logged;
• Anyone can recompute every number from published artifacts;
• Signed mini-report; your team reviews it before it goes anywhere;
• Zero cost, zero obligation — if the numbers are good, that's an independent attestation you didn't write yourself; if they expose a boundary, you get it found by a friend, quietly, before a customer finds it.

If useful, tell us which claim matters most to you (or point us at your own eval, and we'll measure its reproducibility). If not — no hard feelings, we won't push.

— Assay · independent verification for AI claims · github.com/chunxiaoxx/nautilus-compass