akms

A deterministic global-local knowledge compiler. AKMS turns typed knowledge nodes into a directed graph, compiles task-scoped projections from that graph, and ingests structured evidence back into it.

What the core does

• Typed knowledge nodes with a frozen v2 schema, provenance, and trust metadata.
• Global/local ownership — a read-only global vault plus a writable project graph.
• Deterministic selection — tag derivation, subgraph queries, and dedup are pure algorithmic paths. No language model participates in graph operations.
• Projections — task-scoped slices of the graph, reproducible from the same inputs.
• Evidence ingestion — structured outcomes flow back into the graph.

The core performs no network calls and requires no API key.

Installation

pip install akms

That is the minimal core: graph compilation, schema, projections, evidence, and the akms CLI, with four dependencies and no provider SDK.

Optional capabilities are explicit extras:

Extras combine: pip install "akms[learn,failure-memory]". Quote the argument in zsh, which otherwise treats the brackets as a glob.

Consumption modes

AKMS is usable through any of these, and they are parallel supported paths rather than layers you must adopt in order:

• the Python library API
• the akms CLI
• MCP tools
• a skill or agent-host integration
• akms.orchestrator — an optional first-party runtime that drives a coding agent directly
• an external consumer that requests projections and returns evidence

The embedded runtime is one consumer of the same public contracts external systems use. AKMS does not own portfolio-wide orchestration.

Requirements

Python 3.12. Some search paths use qmd, an external Go binary; when it is absent those paths fall back to grep.

Maturity

Research software, published as a preview. See the repository root for limitations, privacy and provider notes, and the stable-versus-experimental capability matrix.

License

Apache-2.0. See LICENSE at the repository root.