ALTK-Evolve publishes code and benchmark pull requests; the ten-point gain comes from the authors' small held-out evaluation slice.
ALTK-Evolve adds learned memory to customer-service agents
2026-10-20Frontier watchTRIAL
IBM Research and Hugging Face report that ALTK-Evolve learns reusable prompt-level guidance from completed customer-service episodes. On a held-out τ²-bench airline set, the authors report a ten-point gain for GPT-5.5 over 20 tasks repeated four times. The Apache-2.0 implementation is public, but the results come from the project team and cover a small benchmark slice rather than live customer traffic.
Availability: Apache-2.0 repository and benchmark pull requests are public
What this means for you
For agent teams, the interesting claim is not another vector store but a memory layer that turns prior outcomes into editable instructions. The public code makes a bounded trial possible; teams should test whether gains survive unseen policies and whether bad episodes poison later behavior.
FRONTIER WATCH
Public artifacts and validated results are different.
The recorded fact is the public work and its artifacts. Author-reported outcomes do not establish independent reproduction. Next checks include usable code, task conditions and third-party validation.
Evidence & related material
Technical details & scores
| Method | learned prompt-level behavioral memory from completed trajectories |
|---|---|
| Evaluation | held-out τ²-bench airline subset; 20 tasks x 4 trials; author-reported 72.5% to 82.5% with GPT-5.5 |
| Availability | Apache-2.0 repository and benchmark pull requests are public |
| Evidence Limit | project-team evaluation; no live deployment or independent reproduction |
| Time Basis | Hugging Face RSS and metadata publish at 2026-10-20T15:19:48Z |
Technical impact 3
Engineering value 4
Adoption signal 2
Maturity 2
Verification cost 3
Risk 3