Memdoor
A multi-agent collaboration project that addresses programming problems by having multiple medium-scale models work in parallel as a team, demonstrating collaboration gains on code benchmarks.
Timeline (last 90 days)
Memdoor is released as open source, and the code and team-solve workflow definition can be viewed directly in the GitHub repository at examples/workflows/team-solve, providing conditions for external reproduction.
Memdoor is an open-source multi-agent orchestration project, and team-solve is one of its example workflows, used to organize multiple models to collaborate on solving programming problems.
The core mechanism of team-solve is to have agents in different roles each take on responsibilities such as problem decomposition, solving and verification, and to pass context among agents through Memdoor's memory layer.
The Memdoor-related post received 243 points and 115 comments on Hacker News, with discussion focused on whether multi-agent systems can reliably surpass a single strong model.
Questions in the Hacker News comment section mainly centered on reproducibility and generalization ability, and on whether the extra latency and engineering complexity of multi-agent workflows are worth it.
The on-machine implementation of "a team of 5 models" is very likely not five complete copies performing inference simultaneously, but rather time-sharing or multiple experts after parameter-efficient fine-tuning.
Memdoor's post did not disclose the specific hardware configuration, inference cost, model division of labor, answer aggregation method, or the specific score on LiveCodeBench Hard.
Memdoor is currently more of an "existence proof" than a reproducible solution.
The Memdoor project used a team of 5 Qwen3.8-27B models and beat Fable 5 on LiveCodeBench Hard.
All Facts (9)
The Memdoor project used a team of 5 Qwen3.8-27B models and beat Fable 5 on LiveCodeBench Hard.
60%UnverifiedMemdoor is an open-source multi-agent orchestration project, and team-solve is one of its example workflows, used to organize multiple models to collaborate on solving programming problems.
50%UnverifiedThe core mechanism of team-solve is to have agents in different roles each take on responsibilities such as problem decomposition, solving and verification, and to pass context among agents through Memdoor's memory layer.
50%UnverifiedQuestions in the Hacker News comment section mainly centered on reproducibility and generalization ability, and on whether the extra latency and engineering complexity of multi-agent workflows are worth it.
50%UnverifiedMemdoor's post did not disclose the specific hardware configuration, inference cost, model division of labor, answer aggregation method, or the specific score on LiveCodeBench Hard.
50%UnverifiedMemdoor is currently more of an "existence proof" than a reproducible solution.
50%UnverifiedMemdoor is released as open source, and the code and team-solve workflow definition can be viewed directly in the GitHub repository at examples/workflows/team-solve, providing conditions for external reproduction.
50%UnverifiedThe Memdoor-related post received 243 points and 115 comments on Hacker News, with discussion focused on whether multi-agent systems can reliably surpass a single strong model.
50%UnverifiedThe on-machine implementation of "a team of 5 models" is very likely not five complete copies performing inference simultaneously, but rather time-sharing or multiple experts after parameter-efficient fine-tuning.
50%