Persistence Routing
Where to write what, so a cold session reaches the current state of a project in one read.
The problem it solves
If you work with an AI across many sessions, or across many projects, the binding constraint is not intelligence. It is continuity. A session that has to reconstruct the state of a project from scattered files before it can help is a session that burns its first ten minutes being useless. Persistence Routing is the discipline of writing things down in the right place so that never happens.
The core idea: different kinds of knowledge have different shelf lives and different readers, so they belong in different documents. Put them in the wrong place and they rot, drift, or hide.
The layers
The brain doc is the standard. It holds the architecture, the conventions, the schemas, the philosophy of a project, written at the project's real altitude. It is current-state only. It is what lets a cold session reach the present in one read. It is not a work log, and the fastest way to ruin one is to fill it with dated session notes until the standing facts are buried.
The timeline is the narrative. Past-tense story of how the project got here: origins, pivots, milestones. This is what a cold session skims to understand recent context and what is still load-bearing.
The changelog is the firehose. Maximum detail, every meaningful task, optimized for reconstruction. This is where the dated work actually goes, so the brain doc does not have to carry it.
Memory is the pointer layer. Thin routing entries that say where the content lives. Memory routes; the brain docs hold. Never bloat the pointer layer with the content it points to.
The read order
A cold start reads in priority order: brain doc first for orientation, then the timeline for recent narrative, then the changelog on demand for depth. The whole point of the routing is that this order is enough to make competent decisions without reading everything.
The anti-pattern
The failure mode has a name worth remembering: the sticky brain doc, the one that slowly accumulates work logs until it is half standard and half diary. Every session adds "today I did X," and within a month the load-bearing architecture is drowned in dated noise. The fix is the routing itself: dated work goes to the changelog, narrative goes to the timeline, the brain doc stays current-state. When you are unsure where something goes, ask what reader needs it and on what cadence. The answer tells you the layer.