Coinbase creates AI 'digital teammate' CEEcil: capable of long-term memory, proactive follow-up, and autonomous judgment Odaily Planet Daily News: Coinbase has disclosed the architecture and application of its internal AI system "CEEcil". The system is not a simple chatbot, but is designed as an AI agent similar to a "team member" that can be mentioned in Slack and has long-term memory, proactive follow-up, and judgment abilities. CEEcil retains team context through a three-layer memory mechanism: the backend continuously extracts information such as decisions, blocked items, and responsible persons from authorized channels; Integrate short-term observations into long-term summaries every night; Retrieve historical memory, recent conversations, and real-time data as needed when answering questions. The system also adopts a hierarchical model architecture, where simple queries are prioritized through APIs or knowledge bases, and only complex, multi-step tasks are handed over to stronger AI models to reduce costs and latency. In practical applications, CEEcil can actively join relevant Slack discussions, send emoji responses, and automatically follow up when no one is handling the issue. Coinbase stated that the system has also refused to submit operational documents containing customer identity information to Git and suggested using a desensitized version or compliant storage method, demonstrating its ability to make certain security judgments. In terms of architecture, CEEcil consists of Go services and AI agent runtime, and is connected to memory retrieval, knowledge base, real-time queries, and Slack tools through MCP. Coinbase has also set up real-time "kill switches", call counts and spending limits, audit mechanisms, and requires humans to review and merge the code it generates. Coinbase stated that CEEcil is currently able to assist teams in handling production accidents and answering expert knowledge that would have taken several hours to obtain; In the future, we hope to expand this model to multiple AI agents for specific teams, so that AI can not only "answer questions", but also keep remembering, actively participate in and complete actual work.