Memory Sidecar has become a critical component for agent‑agnostic memory in production AI workflows. Version 3.5.1 marks a deliberate shift from feature velocity to operational maturity, addressing real‑world pain points observed across deployments. This release introduces no new APIs or schemas; instead it hardens the runtime behaviour so you can rely on it under load, network churn, and adversarial conditions.
Three areas received focused attention: connection resilience, caching, and observability. All are configurable via environment variables or mounted configuration files, allowing operators tune them per workload.
Connection Resilience
When the backing memory store (e.g., Redis or a SQL backend) becomes unresponsive, the sidecar now uses a circuit breaker with exponential backoff retries. The circuit opens after a configurable consecutive failure threshold, rejecting requests immediately to prevent cascading timeouts. A half‑open state periodically probes the backend for recovery. Retries include jitter to avoid thundering herd.
environment:
