CoReMem: Cost-Internalizing Memory for Successive Coding Tasks with Prospective Valuation
On ChainSWE, CoReMem raised a 27B coding agent’s resolve rate from 21.0% to 25.7% while reducing steps and token use.
The framework selects memory for successive coding tasks by measuring whether past work helps on the next task, accounting for the cost of writing and reading it. It builds compact memory around repository locations and executable checks, comparing runs with and without individual segments of prior work. Training rewards task resolution and fewer steps, with extra credit assigned to steps that use memory-supplied anchors. Against the untrained actor without memory, the authors report 17% fewer steps and 11% fewer cache-aware tokens, including memory-module calls.
ArXiv · AI/CL/LG's note
The framework selects memory for successive coding tasks by measuring whether past work helps on the next task, accounting for the cost of writing and reading it. It builds compact memory around repository locations and executable checks, comparing runs with and without individual segments of prior work. Training rewards task resolution and fewer steps, with extra credit assigned to steps that use memory-supplied anchors. Against the untrained actor without memory, the authors report 17% fewer steps and 11% fewer cache-aware tokens, including memory-module calls.
ArXiv · AI/CL/LG's note
score 5