CoReMem
A memory-augmented framework for coding agents. Through forward-looking episode valuation and an actor adaptation mechanism, it internalizes memory costs and improves solve rates and efficiency on sequential tasks.
Timeline (last 90 days)
CoReMem uses prospective segment valuation to quantify the actual value of each trajectory segment for the next task, by comparing the performance on the next task under the same inherited repository state with and without that trajectory segment.
CoReMem guides the construction of a compact memory structure organized around anchors, which include repository locations and executable checks.
CoReMem's actor adaptation mechanism estimates on-policy memory value within rollout groups, comparing performance with and without memory.
CoReMem's reward design prioritizes problem-solving rate while penalizing the number of steps in successful runs.
CoReMem adopts provenance-gated value credit, where the additional value term is allocated only to steps identified as having used anchors provided by memory, while every step still retains outcome credit.
On the ChainSWE benchmark, CoReMem raises a 27B-parameter actor model's solve rate from 21.0% to 25.7%.
On the ChainSWE benchmark, compared with an untrained actor model without memory, CoReMem reduces total steps by 17% and cache-aware total tokens by 11%, with memory module invocation overhead already included.
CoReMem's contribution lies in establishing a framework that uses downstream task performance to guide memory design in reverse, making memory a dynamic component that can be valued, selected, and trained.
All Facts (8)
CoReMem uses prospective segment valuation to quantify the actual value of each trajectory segment for the next task, by comparing the performance on the next task under the same inherited repository state with and without that trajectory segment.
50%UnverifiedCoReMem guides the construction of a compact memory structure organized around anchors, which include repository locations and executable checks.
50%UnverifiedCoReMem's actor adaptation mechanism estimates on-policy memory value within rollout groups, comparing performance with and without memory.
50%UnverifiedCoReMem's reward design prioritizes problem-solving rate while penalizing the number of steps in successful runs.
50%UnverifiedCoReMem adopts provenance-gated value credit, where the additional value term is allocated only to steps identified as having used anchors provided by memory, while every step still retains outcome credit.
50%UnverifiedOn the ChainSWE benchmark, CoReMem raises a 27B-parameter actor model's solve rate from 21.0% to 25.7%.
50%UnverifiedOn the ChainSWE benchmark, compared with an untrained actor model without memory, CoReMem reduces total steps by 17% and cache-aware total tokens by 11%, with memory module invocation overhead already included.
50%UnverifiedCoReMem's contribution lies in establishing a framework that uses downstream task performance to guide memory design in reverse, making memory a dynamic component that can be valued, selected, and trained.
50%