AI Coding Agents Keep Relearning the Same Lessons, Costing Teams Time and Money

AI coding agents like Claude Code and Cursor perform well within a single session but lose all accumulated knowledge once that session ends, forcing every new session to rediscover facts the team already knows. This repeated re-derivation wastes tokens, slows engineers down, and occasionally reintroduces previously rejected fixes that caused problems in production. The core problem is that there is no shared, durable memory layer where agents can store and retrieve team-specific context across sessions and machines. A scratch file or chat transcript stays siloed to one engineer's environment and never reaches a teammate's agent. The authors frame this ongoing, invisible drain on productivity as an 'amnesia tax' and argue the bottleneck is not model capability but the absence of persistent shared knowledge for agents to draw on.