claude-mem 13.20–13.21 Adds a Persistent Quota Breaker for the Observer; Offline Installation No Longer Forces OAuth

claude-mem released versions 13.20.0 through 13.21.2 in rapid succession on August 31. The core changes go beyond the installation UI, redefining how the memory observer handles failures and quota state. Previously, after a provider rejected requests because the quota had been exhausted, every new tool event would still launch another generation that was certain to fail. If the rejection message was returned as ordinary assistant prose rather than an exception, the system would not even record it in its health state. Version 13.20 adds a persistent quota breaker: requests are blocked outright during the cooldown period, and only one probe is allowed through when it expires. The probe claim is tied to the generator that acquired it, preventing an older job from incorrectly clearing a newer probe when it finishes and permanently blocking the provider.

Observer conversations now also use bounded generations seeded from existing memories. When a boundary is reached, a new generation receives a summary of the actual session-start context and continues automatically, so the final observation in a session can be persisted without waiting for the next tool call. If the cmem gateway issues a terminal rejection for a previously issued key, the system can switch to the user’s own Anthropic plan and notify the user only once at the beginning of the session. Such expected fallbacks are no longer misreported as service outages.

The account-first installation flow introduced in 13.20 caused a serious regression: the endpoints required for OAuth had not yet been deployed, causing every installation to stop at “Could not start OAuth login.” Even non-interactive installations explicitly configured to use a local provider failed. Version 13.21 changes `--provider claude` to bypass login and cmem.ai entirely, using the existing Anthropic plan directly. Gemini and OpenRouter still require gateway authorization. The installer also no longer hard-codes the trial period as seven days. In SSH or headless environments, it prints the login URL before deciding whether to open a browser.

These fixes are worth noting for developers of agent tools: quota rejections, service failures, and fallback paths must be represented as distinct states, and circuit-breaker state must persist across worker restarts. However, the releases iterated rapidly within a single day, with 13.21.1 and 13.21.2 making further changes to trial and payment flows. Production environments should pin an exact version and independently test multi-session contention, offline startup, and provider switching.