Coinbase Builds AI "Digital Teammate" CEEcil: Long-Term Memory, Proactive Follow-Ups, and Autonomous Judgment

Odaily News - Coinbase has disclosed the architecture and application of its internal AI system "CEEcil." This system is not a simple chatbot but is designed as an AI agent that functions like a "team member." It can be mentioned in Slack, and possesses long-term memory, proactive follow-up, and judgment capabilities.

CEEcil retains team context through a three-tier memory mechanism: in the background, it continuously extracts information such as decisions, blocked items, and owners from authorized channels; each night, it consolidates short-term observations into long-term summaries; and when answering questions, it retrieves historical memories, recent conversations, and real-time data on demand. The system also adopts a tiered model architecture, where simple queries are prioritized for processing via APIs or knowledge bases, and only complex, multi-step tasks are delegated to more powerful AI models, thereby reducing cost and latency.

In practical applications, CEEcil can proactively join relevant Slack discussions, send emoji reactions, and automatically follow up when issues go unaddressed. Coinbase states that the system also once refused to submit operational documents containing customer identity information to Git, suggesting the use of redacted versions or compliant storage methods instead—demonstrating a certain level of security judgment.

Architecturally, CEEcil consists of a Go service and an AI agent runtime, and connects to tools such as memory retrieval, knowledge bases, real-time queries, and Slack via MCP. Coinbase also has set up a real-time "kill switch," call limits and spending caps, audit mechanisms, and requires human review and merging of the code it generates.

Coinbase says CEEcil is now able to help teams handle production incidents and answer expert knowledge that previously would have required waiting hours to obtain. In the future, the company hopes to expand this model into multiple AI agents tailored to specific teams, enabling AI not only to "answer questions" but also to continuously remember, proactively participate, and complete actual work.