AI agents forget everything between sessions. They loop, burn tokens, and repeat the same failures because no individual turn has context about the ones before it. tapes is an open source telemetry proxy that intercepts LLM API calls, builds a content-addressable Merkle DAG of every conversation turn, and publishes events to Kafka with zero additional instrumentation required (no SDK). The agent points its API base URL at the proxy and keeps running. From there, Flink SQL runs continuous anomaly detection against the telemetry stream: stuck loop detection in tumbling windows, token spike detection against rolling averages, and domain-specific pattern matching. When Flink catches something, it publishes an alert to a separate Kafka topic. An alerts consumer writes that back into the agent's session store. The next run reads those signals before it starts.

To test this architecture out, we did what we knew best which was  built and tested this by speedrunning parallel agents evolving their navigation parameters across Gameboy games. The infrastructure around it was observant. The Merkle DAG structure means every conversation turn is content-addressable and forkable, so you can extract clean datasets from months of agent runs without any post-hoc instrumentation. This pattern for speedrunning games is the same for speedrunning streaming issues, and managing anomalies before they become an issue using self healing infrastructure.

This session covers the full pipeline end to end: Tapes proxy to Kafka, Flink SQL anomaly queries, the alerts feedback loop, observational memory across sessions, and the path from raw telemetry to even fine-tuned your internal and local models. Live demo included. All code is open source.