Mahal Systems MemoryArena
Fits agents whose sessions build on each other (a buyer, planner or researcher that returns to an unfinished job). Borrow the interdependent-subtask design where memory errors compound.
What it asks the agent to do
- carry what was learned in one session into the next dependent subtask
- bundle purchases across sessions under a budget or preference constraint
- narrow a product search progressively using earlier results
- plan a group trip while honoring constraints gathered earlier
- chain formal reasoning steps across sessions
What to borrow from it
case templatesandbox design
Caveats
Task count and grading details are not stated on the paper page or README; data license is CC BY 4.0 but the code license was not visible.
Details
| Workloads | memory, shopping, travel, web browsing, planning |
|---|---|
| Interaction | autonomous long horizon |
| Environment | mixed |
| End state checked | No |
| Tools | simulated shopping site, search, travel planner |
| Grading | programmatic check |
| Size | Not published |
| Data | synthetic |
| Personal data | none |
| License | CC-BY-4.0 |
| Harness | own repo |
| Status | active |