CoReMem decreases the steps of a coding agent by 17%

On 12 November 2026, researchers from Yale University, Tsinghua University and Goodvision AI showed CoReMem. CoReMem gives a coding agent a memory of a code repository across many tasks. The memory keeps only data that helps more than the tokens that it uses, and tokens are units of text. On the Qwen3.8-27B model, the resolve rate increased from 21.0% to 25.7%, and the steps decreased 17% on average.

How it works

Data from tasks that are done adds to the input and makes the agent slow. The authors removed each part of the agent data in turn. Only two types of record helped subsequent tasks:

• Code locations: paths, line numbers and symbols.
• Checks: test commands and exit codes.

The training decreases the number of steps that the agent uses.

The numbers

• On Qwen3.8-27B, token use decreased 11%.
• On three open models, the resolve rate for each token increased 38% to 46%.
• On models with no new training, the steps decreased and the resolve rate did not decrease.
• Three other memory tools used more tokens than CoReMem on five of six models.

What is not known

These are results of the authors on one benchmark. The report is a preprint. No external team shows the same results.

Insight: More agent memory is not always better

Databricks tells that its tests show 3 times more correct results on tasks at work with 31% less token use. CoReMem shows that data from tasks that are done adds tokens, and only locations and checks help. This reads as a move from who keeps most to who keeps the correct data, and a memory result must give a token number. Make each vendor tell you the data that its memory keeps.