Docs.
Mneme is an open-source CLI and local MCP server for governing AI coding agents. Start with one check, connect the Decision Index, then go deeper on proposal review, violations, language coverage, and benchmark methodology.
Run your first governed check.
Install the CLI, initialize the decision corpus, then evaluate one proposed task.
$ pipx install "mneme-hq==0.9.0" $ mneme init $ mneme check --memory .mneme/project_memory.json --input prompt.txt --query "storage backend"
mneme init creates .mneme/project_memory.json with an empty, valid decision corpus. To protect existing decisions, it refuses to overwrite an existing corpus unless --force is supplied.
Standard Unix exit codes for CI gates and pre-commit hooks: 0 = pass, 1 = warn, and 2 = fail.
To enforce on Claude Code Edit and Write calls, install the runtime as above and load the plugin from a Mneme checkout (runtime and plugin are separate artifacts, pending marketplace release):
claude --plugin-dir ./integrations/claude-code-pluginTo generate a Cursor rules file (.cursor/rules/mneme.mdc) from active decision records:
mneme cursor generate --memory .mneme/project_memory.json --query "storage backend"Reference surfaces
Mneme HQ's documentation is intentionally narrow. We document the surfaces engineers need to operate Mneme: the CLI, the local MCP server, human decision authority, the violations Mneme catches, language coverage, and benchmark methodology. Everything else lives in the source repository.
How Mneme turns architectural decisions into a structured corpus with deterministic enforcement across four workflow boundaries: guidance, prevent, catch, verify — and why retrieval is not enforcement.
How enforcement works →What Mneme can attach to an engineering decision — identity, applicability, authority, lifecycle, precedence, enforcement, and evidence — and where each is documented. Architecture is the first decision domain.
Decision model →How a Mneme-ready decision becomes a verified deterministic guardrail: mneme protect list, status, validate, activate. Validation is not activation; protection is only reported when canonical assessment observes enforcement.
Commands, flags, and exit codes for mneme, including the 0.9.0 decision-mcp server and decision authority workflow alongside setup, Audit, protection, enforcement, exports, and benchmarks.
Install and operate the local stdio Decision Index server in 0.9.0. Exact tools, input and result shapes, errors, storage, and the authority boundary.
MCP server reference →Review producer proposals, preserve provenance, and explicitly accept or reject them with the human authority CLI without silently activating protection.
Decision proposal workflow →Twelve worked examples across architecture, workflow, security, dependency, and platform governance. Each shows the rule, the AI's offending output, and the structured flag Mneme emits.
Governance violations →Canonical coverage matrix. Tier 1 (Python, TypeScript, JavaScript) and Tier 2 (Go, Java, C#, Rust). Capabilities, limitations, and roadmap per language.
Supported languages →How the v1.1 governance benchmark is measured: layered retrieval and enforcement scoring, structured-output verification, pre-registered thresholds, anti-gaming protocol.
Benchmark methodology →How to read these docs
Use the CLI reference for commands, the MCP server reference for the protocol surface, decision proposals for human review, the protection activation guide for the audit-to-Protected loop, and the benchmark methodology for measurement. For the ideas behind the system, continue to Insights.
One decision boundary across every coding surface.
Connect Mneme where tasks enter, where agents edit, and where changes merge.
Explore all integrations →