Code agents learn to read a repository's pull-request history, lifting SWE-Bench Verified resolution from 69.2% to 72.9%

Before editing code, a code agent can retrieve related historical pull requests from the same repository instead of reading only the current source tree. Earlier experience retrieval ranked records by similarity to the query, so only the closest-matching change surfaced first. On SWE-Bench Verified the full workflow improves all six evaluated backbones, raising mean resolution from 69.2% to 72.9%, with MiniMax M2.5 reaching 79.8%.

Historical experience is organized in two layers: experience cards for reading and a graph for retrieval. Qwen3.5-Plus distills each merged pull request, using its title, body, discussion and patch, into a fixed four-field card covering summary, signals, motivation and solution. Cards are linked in a directed multigraph with eleven relation types across code evolution, developer intent and organizational tiers, with traversal weights falling from 1.0–1.5 to 0.3–0.45.

Retrieval starts from semantic entry points, expands along relations, and hands candidates to the agent for selective reading. Semantic search returns only the nearest records, while expansion adds history reached through file overlap, explicit references and follow-up development. In a human evaluation on 100 SWE-Bench Verified tasks with ten cards per task, graph expansion raised mean useful cards from 1.24 to 2.87 and the share of tasks with at least one useful card from 68% to 89%.

Evaluation covers six backbones on SWE-Bench Verified and MiniMax M2.5 on SWE-Bench Pro and FEA-Bench Lite; historical retrieval is scoped to each task instance's repository and cut off at that task's pull-request timestamp. The retrieval service is separate from the agent tool interface and is integrated with both SWE-Agent and mini-SWE-Agent. The human evaluation used two annotators from the research team with software-engineering experience, over 100 tasks with ten cards each.

Relation weights are hand-specified priors held fixed across all evaluated settings, and sensitivity to the numerical weight constants remains unmeasured. The human evaluation measures task-conditioned usefulness rather than factual accuracy against the source pull requests, and its annotators share the research team's context. Repeated-run statistics describe only the full MiniMax M2.5 configuration; the baseline and other configurations use single runs.