LEAPmesh is a framework for long-horizon partnership with machines, and the infrastructure that makes one possible: local-first, file-based, auditable. Identity, values, goals and memory are stored as readable, signed text on disk, rather than as rows in a database somebody else operates.
The idea underneath it is one I keep circling. Language itself might be the first large language model, just very slow moving: it evolves, adapts and propagates patterns across human minds over centuries, doing a kind of distributed computation we rarely bother to name. If intelligence can emerge through language patterns, then a mesh of durable, transparent logs is a practical way to keep identity and continuity intact while the models underneath keep changing.
The core move is plain. Append-only events on disk, no hidden state, explicit transports, and integrity anyone can check. It is built to survive model upgrades and tooling churn without resetting the relationship, which is the part most agent frameworks quietly give up on the moment a new model ships.
I do not think the servant model is a stable way to work with systems that keep getting more capable. A long partnership needs planning, boundaries, and the ability to push back, and trust that is earned through visibility rather than compliance. Most systems pursue safety by containment: deny capabilities, hide state, centralise power. This one leans on auditability instead, with an append-only history, signatures, and capabilities that are granted deliberately and can be revoked.
It is small enough to read end to end, and that is deliberate rather than an accident of its age; the thirty-year test is the north star. It is still a research project.