Ai

Claude Mem v12.4.7 Tightens Isolation, Validation and Worker Reliability

Claude Mem v12.4.7 Tightens Isolation, Validation and Worker Reliability

Claude Mem v12.4.7 is a stability-focused release that finishes a broad cleanup of bug-prone fallback behavior. The update replaces silent tolerances and duplicate recovery logic with stricter boundaries, clearer failure modes, and more consistent filtering across services. For teams using the project in multi-account or automated environments, this version is especially important because it improves isolation, reproducibility, subprocess safety, and worker lifecycle handling.

What Changed

The core theme of v12.4.7 is removing two classes of anti-patterns: “defenders,” such as orphan cleanup and duplicate liveness probes, and “tolerators,” such as silent JSON drops and drifted SSE or SQL filters. In their place, the release introduces fail-fast behavior, single-source logic paths, and stricter schema boundaries.

A major infrastructure improvement is stronger multi-account isolation. The project now supports account separation through CLAUDE_MEM_DATA_DIR and a per-UID worker port formula of 37700 + uid % 100, with CLAUDE_MEM_WORKER_PORT available as an override. That reduces the risk of process and data collisions in shared environments.

The release also introduces a clearer internal trust boundary with CLAUDE_MEM_INTERNAL=1, replacing prior observer-session detection based on the current working directory. This is a meaningful architectural cleanup because explicit trust signals are more reliable than environment inference.

Consistency was improved through a shared shouldEmitProjectRow predicate that keeps SSE broadcast behavior aligned with pagination filters. The package stack was also tightened by pinning chroma-mcp to version 0.2.6 for reproducible installs.

Operationally, install and uninstall flows now use a shared shutdown-helper to release file locks before overwriting or deleting files. Database migration 30 adds an observations.metadata column, and spawned subprocesses now have proxy environment variables stripped so user proxy settings do not leak into AI API calls.

The post-review fixes add another layer of hardening. Worker-service restarts now exit with code 1 if spawnDaemon fails, rather than failing ambiguously. The shutdown helper now distinguishes between AbortError from a slow worker and a connection-refused state from a worker that is already gone. Additional fixes include proper quoting for $HOME cache lookups in hooks.json, Windows compatibility for the timeline-report skill without requiring process.getuid(), and validation fixes in the OpenCode plugin path.

Why It Matters

This version matters because it reduces hidden failure modes that are often difficult to diagnose in production. By removing silent tolerance patterns and replacing them with explicit checks and shared predicates, the release should make behavior more predictable across worker management, event streaming, database filtering, and plugin boundaries.

The multi-account and trust-boundary changes are particularly relevant for AI tooling operators, development teams, and enterprises running shared machines or multiple user contexts. Better isolation and explicit internal signaling can reduce cross-account interference and make deployments safer to automate.

Just as importantly, the review-comment fixes show the maintainers are tightening operational edge cases rather than stopping at structural refactors. That makes v12.4.7 less about new features and more about reliability engineering, which is often what mature AI infrastructure projects need most.

Official Source: https://github.com/thedotmack/claude-mem/releases/tag/v12.4.7