Skip to content

Modes

Artesian is non-intrusive: set it up once and keep driving your agent exactly as before. You opt into only the capabilities you want. Mode is a config choice (artesian init writes artesian.toml).

memory (default, zero workflow change)

Exposes durable memory to your agent over MCP (memory.find / memory.store) plus a session- start context injection. You do not change how you prompt or run your agent — you simply gain faster, cheaper context because relevant knowledge is retrieved instead of re-read. No orchestration, no sandbox. memory.backend selects files (Open Knowledge Format md bundle, zero infra), sqlite-vec (local hybrid, zero infra), or feature-gated qdrant (shared server). See memory.md.

orchestrate (opt-in)

Adds the optional master/worker/judge roles and the headrace task queue. Composable: master-judge only, one agent bound to all roles (e.g. Codex everywhere), or the full triad. No mandatory agent loop — leave delegation off if you don't want it. See task-tracking.md.

full

memory + orchestrate + the optional sandbox Docker sandbox for isolated workers.

advanced (bring-your-own)

For power users who already have a memory layout. Point Artesian at an existing markdown tree (including any OKF bundle) or vector collection and it adapts and overlays without owning or rewriting your schema: it reads your structure, serves retrieval over it, and adds Artesian capabilities on top. You keep full control of your data model; Artesian meets it where it is.

Loop memory kit (portable across agents)

The loop memory kit is the anchor-set bundle that makes Artesian portable across Codex and Claude Code (or any MCP-compatible agent). It lives in .artesian/kit/ and contains:

  • vision.md — project purpose, goals, current phase.
  • agents.md — agent roster, roles, capabilities.
  • kit/index.md — manifest + session-start recipe.
# Create the kit (once per project):
artesian kit init --vision "Build a memory control plane for agent loops"

# At session start (any agent, any model):
artesian memory anchor recover    # restore last anchor + targeted recall
artesian kit status               # print vision + anchor summary

# Export the full kit as a single markdown bundle:
artesian kit export               # stdout
artesian kit export --output kit-bundle.md

Over MCP the same kit is available as: - memory.kit.get — returns vision, agents.md, and last anchor in one call (load at session start). - memory.kit.set — updates the vision (call when the phase changes).

The kit works identically in Codex and Claude Code. Swap the model; keep the kit. This is the "own your learning loop" principle: the company-veteran context travels with the project, not with the agent.

Switching modes never requires re-architecting your project. Start in memory mode for the token win; add orchestrate or advanced later if and when you want them.