Knowledge Accumulates, State Supersedes: Structuring and Carrying Context for Multi-Session, Multi-Harness AI Agent Work — A Design Study of The Curator, an Open-Source Context Engine

Large language model (LLM) agents now carry out work that outlasts a single session: building software, running literature reviews, writing books, producing campaigns. Every session, however, begins with an empty context window, and when the work moves to a different agent harness, a different model or a different computer, whatever the previous session knew is lost. This article argues that the binding constraint in such work is the session boundary rather than the size of the context window, and that the remedy is to treat context as owner-controlled infrastructure with an explicit structure. I propose a model of context for agent work that distinguishes three kinds by how they must behave when written to: compounded knowledge, which accumulates; volatile working state, which supersedes; and canonical documents, which are replaced whole and read verbatim. I then describe The Curator, an open-source, local-Girst application built in public over six months and more than 200 releases, which implements this model as plain markdown Giles in a folder the user owns, synchronised through the user's own private repository and read and written by any agent harness through the Model Context Protocol. The system spans three audiences-the individual's second brain, a team's shared brain, and the agents' working context-within one store. I report small, dated measurements: carrying working state moved an agent's identiGication of the correct next priority from 0 of 4 to 8 of 8 runs, and suppressed proposals the project had already ruled out; the reliability of saving depended far more on where an instruction was placed than on how it was worded; and a handoff cured ignorance but not disagreement. The article closes with limitations and open questions for a Gield in which context has become the principal determinant of output quality.