Session Memory for AI Coding Assistants: Automatic Capture, Semantic Retrieval, and Cross- Session Context Preservation

AI coding assistants lose context at three points: session boundaries (each new session starts blank), in-session context truncation (earlier turns are silently dropped or summarized when the context window fills), and developer handoffs (existing mechanisms for sharing AI-assisted reasoning across team members are manual and incomplete). We present a framework that addresses all three failure modes through a combination of lifecycle hookdriven automatic capture, per-repository local storage, and model-directed semantic retrieval. The system uses a session-termination hook to read the on-disk session transcript — bypassing in-memory context loss — extract clean developer-assistant turn pairs, and persist them to a git-trackable local store scoped to the working directory. An MCP-registered retrieval tool backed by a hybrid FAISS+BM25 vector index with Reciprocal Rank Fusion enables the AI assistant to autonomously recover prior context on demand, without explicit developer prompting. The framework is local-first, requires no cloud dependency, and enables team memory inheritance via standard git clone.