Detailed Analysis
A Mac Studio user with 64GB of RAM and 4TB of storage encountered an unexpected system-level out-of-memory error while running Claude alongside other applications, with the system reporting Claude consuming approximately 897MB and Apple's Pages application consuming a staggering 231GB of virtual memory. The error caused the operating system to pause multiple applications simultaneously — a behavior macOS triggers when physical RAM and swap space are both exhausted. The scale of the Pages figure strongly suggests the 231GB reading reflects virtual memory or swap allocation rather than physical RAM usage, a distinction macOS's Activity Monitor can obscure for users unfamiliar with how the operating system reports memory pressure. Claude's 897MB footprint, by contrast, falls within a plausible range for a browser-based or Electron-wrapped application running an active session with accumulated context.
The incident connects to a documented pattern of memory accumulation in Claude-related tooling. Claude Code, Anthropic's agentic coding application, has been reported to cause JavaScript heap overflow crashes when Node.js exceeds its approximately 2GB heap ceiling — particularly during file-listing operations on large directory structures. Separately, Claude's persistent session history architecture stores complete records of conversations, tool calls, and agent interactions in `.jsonl` files, and at least one documented case involved a user whose accumulated Claude session history reached 4.6GB, directly triggering out-of-memory failures on their system. While the Reddit user's situation appears to involve the standard Claude interface rather than Claude Code specifically, the underlying dynamic — gradual, unnoticed accumulation of memory resources tied to Claude sessions — is consistent across both contexts.
The broader architectural picture adds meaningful context. Claude Code's source code, briefly exposed in March 2026, revealed a three-layer memory architecture consisting of a `memory.md` file for manual notes, grep-based search functionality, and an unshipped background daemon component codenamed Chyros. This design prioritizes persistent recall of past interactions and work sessions, enabling Claude to reference prior conversations and tool invocations across sessions. However, that same persistence model means memory footprints can grow substantially over time without explicit user management, creating conditions where long-time or heavy users are disproportionately likely to encounter resource exhaustion events.
The incident illustrates a tension inherent in deploying AI assistants as persistent, stateful applications on consumer hardware. As AI tools move beyond stateless query-response interactions toward maintaining rich session histories and background daemons, their resource profiles begin to resemble those of databases or development environments rather than simple chat clients. The 64GB RAM configuration of the Mac Studio involved is well above average consumer specifications, making it notable that even this relatively high-end machine encountered memory pressure — suggesting the issue is driven by software behavior rather than hardware inadequacy. Practical mitigations available to affected users include clearing old Claude session logs stored in `~/.claude/projects`, monitoring swap usage separately from RAM in Activity Monitor, and restarting Claude periodically to release accumulated heap allocations.
Read original article →