This is a vision / position paper, not an empirical systems report. State-of-the-art AI coding assistants are stateless across development sessions: when a workspace session terminates or a chat thread is cleared, the interactive transcripts that produced design rationales, rejected dependencies, and debugging insights are discarded. Traditional Code Retrieval-Augmented Generation (RAG) frameworks index the static state of a codebase (what it contains), but remain blind to the developmental history (what was learned) throughout developer–AI collaborations. We propose Interaction Context Maps (ICM) as a direction for treating developer–AI interactions as durable, repository-aware memory. The note sketches two ideas we believe deserve deeper exploration: a software-specific typed memory taxonomy, and dual-track staleness checks that bind stored context to concrete syntax trees as the repository evolves. We outline journaled capture, task-intent retrieval gating, branch-scoped isolation, and a lifecycle for unwritten code intents, and we close with open questions and a future evaluation agenda. No implementation results are claimed.
Ruturaj Panditrao (Fri,) studied this question.