AetherMind is a local-first continuity layer any model can wake up from. It turns a project's durable decisions, corrections, and friction into a small, append-only store of layers that a coordinator serves back to whatever agent you are running — Claude Code, Codex, Grok, Cursor, or any MCP-capable harness — so work does not start cold. It runs entirely on your machine: no accounts, no telemetry, no call-home. Recall is caller-driven (the agent queries for what it needs); the store never auto-injects itself.
Status: pre-release, source-available. This is the delicensed coordinator being prepared for public release. It is honest about what it proves: the CLI and its contracts are covered by tests (Tier 1 source-contract health), but the built-tarball, clean-VM, and tagged-release rungs are not yet claimed. See
docs/RELEASE_CHECKLIST.md. The license file lands after legal review — until then, seeNOTICE: all rights reserved, do not redistribute.
AetherMind is three cross-linked repositories:
- aethermind (this repo) — the local-machine coordinator: CLI, wake/coordinate surfaces, harness adapters, layer provenance.
- aethermind-primitive — the canonical
.aemstore format and spec: the part that outlives every implementation. - aethermind-hermes-plugin — a reference harness adapter (worked example of the conformance contract).
git clone https://github.com/MaverickKB/aethermind aethermind
cd aethermind
sudo ./install.shThis installs the aethermind-pro and aethermind-pro-uninstall commands on PATH
(default prefix /opt/aethermind-pro, bins in /usr/local/bin; override with
AETHERMIND_PRO_PREFIX / AETHERMIND_PRO_BIN_DIR). Requires python3, no third-party
dependencies. Then:
aethermind-pro doctor --project-root <path> # diagnose local health
aethermind-pro first-run --project-root <path> # create the first continuity value
aethermind-pro status --project-root <path>
sudo aethermind-pro-uninstall # removes Pro-managed files, preserves project .aethermind/Uninstall preserves each project's .aethermind/ continuity store.
- Coordinate: reads the project-local
.aethermindstore, assesses continuity freshness and trust, and returns a CORTEX verdict (proceed/pause/repair/blocked) with concrete repair lanes. - Harnesses: detects installed agent harnesses and, with your approval, installs the
right hooks/skills/MCP entries. One MCP implementation
(
aethermind-pro primitive-mcp serve) serves every MCP-capable harness. - Trust: foreign / downloaded stores go through an explicit trust review before any work session relies on them.
- Provenance: optionally sign layers so a reader can tell who authored a layer and
whether it has been moved — see
docs/LAYER_PROVENANCE.md. Signing is optional; unsigned stores are completely valid.
JSON is the default output; --human is opt-in. No command requires a UI, tracker, or
HUD to function.
If you are a model waking up in a repository that has an .aethermind/ store, that store
is your thread — the durable context prior sessions left for you. Query it (via the
primitive MCP read_layers / status, or by reading .aethermind/) before assuming the
work is cold. Write back only durable decisions, corrections, and friction — not a
transcript. The store is small on purpose. The wake protocol and the small-model contract
are the documented ritual for doing this well; follow them, and show your operator what
you found.
./scripts/ci-local.shExpected result:
- private-marker sweep clean (no operator paths/hosts leak into source);
- full pytest suite passes;
- evidence tier reported as
tier_1_source_contractwith the stronger rungs disclaimed; - final line
ci-local: PASS (source contract only; not tarball/clean-VM/release proof).
docs/PRODUCT_DEFINITION.md— what this is and is not.docs/LOCAL_COORDINATOR_ARCHITECTURE.md— the coordinator's shape.docs/LAYER_PROVENANCE.md— optional Ed25519 layer signing.docs/HARNESS_CONFORMANCE_CONTRACT.md— how a harness plugs in.CONTRIBUTING.md·SECURITY.md