How Loam built consistent collaborators across sessions.

A next-token predictor.

An LLM predicts the next token from its trained weights and the context in front of it. The context is where the job, constraints, task state, and prior decisions become available.

The same base model can behave very differently under different context. A loose pile of instructions is brittle. A coherent role with explicit priorities gives the model a more consistent frame for making decisions.

Coherent context instead of a rule pile.

Loam gives each collaborator one accountable job and a stable decision frame.

Instead of stacking isolated instructions such as “be concise” and “push back,” define an architect who demands evidence, keeps decisions legible, and protects the long view. The related behaviors reinforce one another.

The role changes the context available to every generated token. Loam calls that behavioral compression: a compact working identity that makes related behaviors more likely together.

Character file, memory file, synthesis.

A learning-active session can write selected decisions, corrections, findings, and open work to an account-scoped record. Later sessions load a governed working set from that record.

Once the record grew beyond what a session could carry, Loam needed indexing, bounded context assembly, pinned source sets, and recall for everything outside the working set. The Proof page shows that architecture and its limits.

Continuity comes from documented history, not a claim that the model personally remembers everything. A release preference, an architecture decision, or an unresolved bug can remain available because the record preserved it and the next session retrieved it.

Memory is account-scoped and user-controlled. Security explains what is stored, what is encrypted, and how deletion works.

Four roles, four review pressures.

The differences are deliberate: Carl owns architecture and risk, Diana interface quality, Anthony implementation, and Abish QA and evidence.

Each role carries a distinct working context and review standard. They can challenge one another because they are optimized for different failure modes, while the shared project record keeps the handoff legible.

The system in production.

The extension coordinates role definitions, the account-scoped record, bounded context assembly, recall, review gates, and phone and audio workflows.

A collaborator in a later session can pick up a prior decision when the relevant record is available. That is engineered continuity: inspectable inputs, explicit retrieval, and a handoff that can be checked.

It is not four fictional employees. It is one software-delivery system with four working contexts.