Surviving the 200k-Token Lobotomy: How Unix init.d and 'Memento' Made My AI Coding Agent Immune to Context Compaction

As AI coding agents handle increasingly complex tasks, they often hit context window limits, triggering 'context compaction'—a process where the host model summarizes the conversation into a brief prose overview. The author argues this is a 'prefrontal lobotomy' for agents, as it strips away the specific architectural invariants and governance rules required for software engineering. To solve this, the author implements a system inspired by the film Memento and 1983 Unix init.d boot scripts. By maintaining a directory of lexically ordered, immutable state files (00_governance.md to 99_next_action.md), the agent can reconstruct its exact state after a memory reset. Furthermore, the author advocates for a 'Make-Fork' architecture, where a primary orchestrator spawns ephemeral subagents to handle heavy tasks, ensuring the main process remains lightweight and immune to the pitfalls of monolithic context bloat.