This experiment tested Claude Code's ability to perform real repository tasks with a memory layer. Unlike previous tests that focused on answer correctness, this experiment required the agent to modify files and pass deterministic checks. Claude Code with the RE-call memory system achieved a 58.3% success rate, significantly outperforming the bare Claude Code at 50.0% and its own CLAUDE.md baseline at 36.1%. The RE-call system's improvement was most pronounced on tasks sensitive to project-specific memory. Interestingly, the static CLAUDE.md file performed worse than the bare configuration, suggesting static instructions can introduce noise. RE-call demonstrated its utility by only retrieving relevant context when needed, rather than simply adding more tokens. The memory layer was consulted in most eligible sessions and provided useful context. While RE-call increased token usage and cost, it aimed to prevent expensive failures from outdated information. The experiment also encountered operational limitations with GPT-5.3 Codex, preventing a definitive comparison. Ultimately, this validated that production memory layers can improve the success rate of agents on real repository tasks by retrieving crucial project-specific context. Future work will focus on optimizing memory retrieval for efficiency and accuracy.