Skip to main content

Context & durable memory

As a project grows, the single-context approach hits a wall. One model trying to hold the whole project in one window runs into two problems at once:

  • Cost. A large context is expensive on every turn.
  • Dilution. A model attends worse as its window fills — the signal you care about competes with everything else loaded in. And the usual escape hatch, compaction, is lossy: it throws away detail to make room.

lmctl's answer is division of context plus offline durable memory.

Division of context​

Split the work so no single player has to hold everything:

  • The Lead focuses on delegation — planning, assigning jobs, and retaining the whole-project picture. It carries the map, not every street.
  • Each Coder focuses on one task with a tight, task-scoped context. It needs to know its job and the code it touches, not the entire history.

This keeps every window small enough to stay sharp, and the cost where it pays off.

Refresh instead of compact​

When a Coder drifts — its context gets muddy, or it wanders off the task — lmctl's move is not lossy compaction. The Lead simply refreshes it: starts a fresh session for that member. Because the canonical knowledge lives outside the session (see below), a fresh session loses nothing important. You get a clean, sharp context back instead of a compressed, degraded one.

durable-memory: the shared brain​

durable-memory is the provider-agnostic record of the project, stored with the project as markdown. Every member reads it, whatever provider backs that member. It is the canonical project knowledge layer; provider sessions are just disposable caches.

Because the canonical state lives in durable-memory, it survives:

  • fresh sessions — a refreshed Coder reads back in,
  • provider swaps — switch a member from one provider to another and the knowledge carries over,
  • session corruption — a broken native session is no longer a loss,
  • project relocation — durable-memory travels with the project directory.

The net effect​

Division of context plus durable-memory means project size is scalable, runs can span days or weeks, and the team's overall memory stays sharp the whole time — without paying for one enormous, diluted context window.

Backed by real long-running projects. A context-budget diagram and a multi-week case study are being prepared and will be presented here.