Codex Squashes Memory Leaks in Long Turns

This is an unglamorous but essential upgrade: long-running agents are only useful when they remain stable under sustained workloads.

• –Codex’s `/goal` feature depends on persistent, multi-turn execution, so memory growth directly undermines its core value.
• –Earlier reports documented runaway memory from parallel agents, retained event buffers, and oversized diff tracking.[GitHub issue #23749](https://github.com/openai/codex/issues/23749) [GitHub issue #39231](https://github.com/openai/codex/issues/39231)
• –Developers should still validate the fix across large repositories, subagents, MCP tools, and multi-hour sessions.
• –Resource hygiene is becoming a competitive feature for coding agents, alongside model quality and tool access.
• –The claim comes from the Codex team rather than a published benchmark, so real-world stability data will matter next.