Developers are increasingly running fleets of coding agents side‑by‑side on a single workstation. Traditional terminal multiplexers such as tmux and newer agent managers focus on window layout, not on memory governance. When ten agents each launch language servers, test runners and browsers, existing tools cannot answer: how much memory belongs to each agent, whether a terminated agent's descendants are fully reclaimed, whether any child process has escaped its agent, or how to keep the system from hitting the swap cliff under over‑commit.
We treat these as runtime‑verification problems: the substrate hosting agents must continuously emit observable signals that operators or auditors can check while agents run.
MemMux is a local runtime that turns resource governance into checkable signals – per‑agent memory attribution, complete reclamation under pressure, visibility of escaped children, bounded footprint when over‑committed, and monitoring overhead. We benchmarked MemMux against tmux, a purpose‑built agent multiplexer and a raw‑process baseline using identical workloads.
With a 7.5 GiB memory budget on a Linux host, MemMux admits a subset of agents and reclaims resources under pressure, keeping the fleet within budget and producing zero swap. Ungoverned tools run every agent, double the RAM usage and spill about 2 GiB into swap. MemMux reclaims 100% of a terminated agent's process subtree, while the raw baseline only reclaims roughly half. It also detects all escaped children (10/10).
The cost of a 1 Hz attribution scan is about 0.6% CPU for one agent and 2.7% for ten agents, slightly above our 2% target. Running the harness on real Claude Code sessions still yields 100% attribution accuracy and low overhead. We release the engine, benchmark suite and a one‑command reproducer.
Review: MemMux offers practical runtime visibility and resource governance for parallel coding agents, achieving full reclamation and escape detection with modest overhead, thus addressing a gap left by existing multiplexers.