The Memdoor project introduces a multi-agent harness designed to solve complex programming problems by coordinating five distinct AI roles: a planner, two ideators, a solver, and a manager, supported by a repair worker. Unlike many agentic systems that rely on long chat histories, this workflow requires each agent to start from a fresh context and communicate exclusively through files on disk. This 'ledger discipline' is intended to maintain clarity and prevent the degradation of reasoning that often occurs in long-context windows. The system was tested using the Qwen3.8-27B model, which is a 27-billion-parameter large language model (LLM) developed by Alibaba.
The Memdoor project introduces a multi-agent harness designed to solve complex programming problems by coordinating five distinct AI roles: a planner, two ideators, a solver, and a manager, supported by a repair worker.
In a series of tests on the LiveCodeBench Hard benchmark—a set of difficult coding problems—the Memdoor workflow achieved a mean score of 9.8 out of 10 over five passes on the first ten problems. This performance matches or exceeds the results of Fable 5, a high-performing coding agent, which scored an 8.0 on the same set. The results were achieved by rebuilding a specific multi-agent method (GVS5H) into the Memdoor framework. The system operates by moving through a directed acyclic graph (DAG) of steps: the planner creates a task plan, ideators propose different algorithmic approaches (e.g., a direct provable algorithm versus a different technique), the solver implements the chosen path, and a review/repair cycle corrects errors based on actual test outputs rather than model claims.
While the results are promising, the project highlights specific limitations and technical trade-offs in multi-agent coordination. For instance, the manager agent occasionally failed to recognize when a program was already correct, leading to unnecessary repair rounds or overwriting successful code. Additionally, the study notes that because the benchmarks date from early 2025, there is a possibility of data contamination, where the model may have seen the problems during its training phase. Despite this, the Memdoor framework provides a reproducible method for achieving high-tier coding intelligence using relatively small, open-weights models by optimizing how those models interact and share state.