Maybe Coding Agents Don’t Need a Bigger Memory. Maybe They Need Continuity.

Coding agents repeatedly lose their operational thread between sessions because they rely on memory abstractions (large context windows, vector stores, chat history) that don't preserve structured execution state. The key distinction is between context (what's available now) and continuity (what lets the next run resume from what actually happened). A better approach stores repo-local artifacts — active work state, failure memory, handoffs, validation expectations — with a simple resume/work/finalize lifecycle. This keeps continuity small, inspectable, evidence-weighted, and tied to the repository rather than a provider-specific black box. The architecture also covers execution contracts, guardrails at action boundaries, and MCP integration, while acknowledging the overhead tradeoffs and when continuity pays off (multi-session, cross-agent, large repos).