Installed per repo, but stateful per machine — one vault and one on-host config serve every project, with a per-repo escape hatch when a machine has no vault.

The harness installs into each project, but the memory it reads and writes is device-wide: a single vault and a single on-host config file (.agentm-config.json) serve every repo on the machine. That is what lets a lesson learned in one project surface in the next, and why "where does state live" is answered once per device rather than once per repo.

Two device-level artifacts plus a per-repo override decide where any phase write lands:

Artifact	Scope	Decides
.agentm-config.json	device	the default state_mode (vault or local) + vault path
<vault>/	device	the durable store, shared across all projects
<repo>/.harness/.project-mode	per-repo	overrides the device default for one repo

The default is vault mode: state routes through the device vault. A machine with no Obsidian/Drive vault opts into local mode (install.sh --local-state), and every phase write lands in <repo>/.harness/ instead — no vault required. The per-repo .project-mode marker wins over the device default, so a single repo can diverge without flipping the machine.

• AgentMemory — the vault this layer makes device-wide. The substrate is where; AgentMemory is what.
• Phases — every phase read/write resolves through this layer to either the vault or the repo-local directory.
• Host adapters — each host reads the same on-host config, so state resolution is identical wherever the harness runs.

Detail:

• Installer CLI — the --local-state flag and the agentm_config.py state-mode setter.
• Project config · Repo layout — the config fields and where the .project-mode marker sits.
• Single-repo state mode — the vault-vs-local resolution model.
• Run without a vault — the recipe for opting a machine into local mode.

Designs:

• Device-Wide Architecture — the full model behind the device-wide substrate.

Architecture · Designs · Home