Mahal Systems LoCoMo-Plus
Fits assistants where remembered preferences or constraints must shape later answers even when the user does not repeat them. Borrow the cue-then-trigger case design and the correct/partial/wrong judge rubric.
What it asks the agent to do
- apply an implicit constraint the user stated earlier (a goal, a state, a value) to a later request that does not mention it
- keep an answer consistent with a cue buried in a long conversation
- recognize when a later query is semantically unrelated to the earlier cue but still bound by it
What to borrow from it
case templategrading design
Caveats
Built by stitching generated cue dialogues into LoCoMo conversations, so it inherits LoCoMo's small base; sample count and repo license were not stated.
Details
| Workloads | memory, knowledge qa, general assistant |
|---|---|
| Interaction | single turn |
| Environment | static |
| End state checked | No |
| Tools | none |
| Grading | llm judge |
| Size | Not published |
| Data | synthetic |
| Personal data | none |
| License | unknown |
| Harness | own repo |
| Status | active |