ACI Personal Agents improved personalized behavior over the base profile, reproduced that behavior exactly after restore, and returned cleanly to the base profile after tagged forget.
0.7812
Personalized score
Local profile active
+31.6%
Lift over base
0.7812 vs 0.5938
0.7812
Restored score
Snapshot parity
24
Tagged items removed
Forget-by-tag
April 2026
Desktop and device-local agents are moving from demos into daily software. Teams building OpenClaw, Moldbot, and similar local agent products do not just need a larger model. They need local memory that can be added per user, restored exactly, and removed cleanly when a profile changes or a user resets a device.
This benchmark tests that product contract directly. It asks four buyer-visible questions: does the local personalized profile outperform the base profile, does restore reproduce that personalized behavior, does tagged forget remove the local contribution, and does the resulting profile return to the base state afterward.
The personalized local profile outperformed the base profile on the benchmark task while keeping the change local to the device-side state.
The restored profile reproduced the personalized score exactly, showing that the saved local state can be reloaded without drift in the measured behavior.
After tagged forget, behavior returned to the base profile. The local profile can move forward and backward without requiring a cloud-side retraining cycle.
The base profile scored 0.5938. The personalized local profile scored 0.7812. Restore reproduced the same 0.7812 behavior, and tagged forget returned the profile to 0.5938 after 24 local items were removed.
| Profile state | Score | What it shows |
|---|---|---|
| Base profile | 0.5938 | Shared global profile without local user-specific additions. |
| Personalized profile | 0.7812 | Local additions improved task behavior while staying device-local. |
| Restored profile | 0.7812 | Snapshot restore reproduced the personalized state without measurable loss. |
| Post-forget profile | 0.5938 | Tagged local memory was removed and behavior returned to the base level. |
What the result says
These results place ACI Personal Agents in the part of the stack where user-specific change has to stay local. For desktop assistants, OpenClaw-style operators, Moldbot-style workflow tools, and device software, the product value is straightforward: personalize on device, restore that state exactly, and remove it cleanly when the profile changes.
The result is directly relevant wherever user-level state has to stay close to the product surface instead of becoming another cloud-retention problem.
For OpenClaw, Moldbot, and similar local or desktop agent products, the benchmark shows the practical contract buyers care about: learn locally, restore locally, and forget locally.
Local operator copilots often need fast profile updates without shipping sensitive behavior or user preferences back to a cloud training loop.
The same benchmark shape matters for laptop, phone, wearable, and embedded personal software where reset and erase must be explicit parts of the product boundary.
Explore where ACI Personal Agents fits in desktop agents, local assistants, and privacy-sensitive device software.