Skip to content

Repository files navigation

Mneme HQ

Architectural drift prevention for the agentic AI SDLC.

Mneme turns architectural decisions and ADRs into deterministic guardrails for the agentic AI SDLC — across coding agents, repository mutations, generated rules, and CI gates.

Tests PyPI Python License: MIT

Mneme is the architectural governance layer behind that drift-prevention mechanism. It keeps recorded engineering decisions active as AI coding systems propose and modify code, instead of leaving ADRs as passive documentation.

Current phase: Layer 1 validation. Retrieval, enforcement, and benchmark semantics are governed by the accepted architecture and freeze record. See Current Phase before changing core behavior.

What Mneme does

Mneme separates architectural guidance from deterministic enforcement:

  • Records architectural decisions in a structured, auditable decision corpus.
  • Retrieves relevant decisions when an agent or model needs architectural guidance.
  • Enforces governed rules deterministically under explicit applicability semantics.
  • Integrates at the earliest reliable boundary exposed by each coding workflow.
  • Audits bypassable mutation paths where pre-change blocking is not technically available.
  • Runs in CI as a final deterministic gate before incompatible changes are accepted.

The same input and governed decision state produce the same enforcement result. Mneme does not depend on an LLM judge for its core allow/warn/fail decisions.

Mneme is not a general-purpose vector store, conversational memory system, autonomous coding agent, or deployment observability platform.

Install

Requires Python 3.11+.

pip install mneme-hq

Verify the CLI:

mneme --help

For repository development:

git clone https://github.com/MnemeHQ/mneme.git
cd mneme
pip install -e ".[dev]"

60-second enforcement example

Initialize a project-local decision corpus:

mneme init

Record one architectural decision:

mneme add_decision \
  --memory .mneme/project_memory.json \
  --id config-format \
  --decision "Use JSON for configuration files" \
  --scope config \
  --constraint "Use JSON only" \
  --anti-pattern "Do not use YAML"

Create a proposed input that violates it:

python -c "import pathlib; pathlib.Path('prompt.txt').write_text('Set up a new YAML config file', encoding='utf-8')"

Run the deterministic check:

mneme check \
  --memory .mneme/project_memory.json \
  --input prompt.txt \
  --query configuration

In strict mode, the prohibited YAML proposal returns a FAIL verdict and exit code 2. A compliant JSON proposal returns PASS and exit code 0.

The CLI is the common enforcement surface. Agent integrations translate their native events into the same Mneme decision and enforcement model.

How it works

Architectural decisions / ADRs
            |
            v
   structured decision corpus
            |
      +-----+--------------------+
      |                          |
      v                          v
relevant guidance       deterministic enforcement
   retrieval             + applicability checks
      |                          |
      +------------+-------------+
                   |
                   v
       workflow-specific boundary
                   |
      +------------+-------------+
      |            |             |
 pre-change     post-change      CI
   hooks          audit          gate

Mneme applies governance at the earliest reliable boundary a workflow exposes:

  1. Before generation when architectural context can be injected into the model call.
  2. Before supported file mutations when an agent exposes a blocking pre-tool hook.
  3. After bypassable mutations through bounded working-tree audits where shell/script writes cannot be inspected safely before execution.
  4. Before merge through CLI-based CI gates.

These boundaries are complementary. An integration only claims the surfaces that have been implemented and validated for that harness.

Retrieval is not enforcement

Decision retrieval answers: which architectural decisions are useful as guidance for this task?

Enforcement answers: does the proposed change violate a governed rule that applies here?

Those concerns are intentionally separated. See ADR-017, ADR-019, and ADR-020.

Supported surfaces

The authoritative support matrix lives in docs/integrations/README.md. The labels below are evidence levels, not interchangeable marketing terms.

Support level Surface
Native integration Claude Code
Native integration Claude Agent SDK
Native integration Google Antigravity
Native integration Codex CLI
Native integration Kiro CLI 3.0 / v3
Validated compatibility Paperclip — CLI and ACP transports, no adapter required
Rules export Cursor
CLI-based CI gate GitHub Actions, GitLab CI
Experimental OpenCode
Planned Deep Agents middleware POC

Each integration documents its actual blocking boundary, bypass paths, degraded behavior, and validation evidence. Start with the integration matrix, not assumptions based on another harness.

ADRs and project memory

Mneme can compile architecture decisions into structured governance records rather than treating ADRs as passive prose.

The repository governance source of truth is .mneme/project_memory.json. The ADR import path preserves explicit source provenance where available so typed rules can be inspected and enforced consistently.

See:

Architecture guarantees

Three principles govern the current mechanism:

  • Deterministic > clever. Enforcement behavior must be reproducible.
  • Auditable > autonomous. A verdict should be traceable to the decision, rule, applicability state, and evidence that produced it.
  • Prevention before review. When a reliable pre-change boundary exists, use it; when it does not, surface the limitation and audit later rather than pretending the path is blocked.

The current Layer 1 scope, frozen surfaces, accepted amendments, experimental work, and deferred Layer 2 work are maintained in docs/architecture/current-phase.md.

Do not infer architecture from this README when a linked ADR or architecture document is more specific.

Benchmark and validation

Mneme's benchmark is a regression and integrity instrument for retrieval and enforcement behavior. It is not a general model-quality benchmark.

The benchmark keeps retrieval and enforcement scoring distinct so changes cannot silently improve one surface while regressing another.

See:

Demos

More examples: mnemehq.com/demo

Contributing

Before changing retrieval, enforcement, applicability, conflict handling, or benchmark semantics, read the architecture and ADRs that govern that surface.

Core behavioral changes may require the repository's charter-amendment procedure. Documentation, tooling, integrations, and examples do not automatically authorize changes to frozen behavior.

License

MIT. See LICENSE.

Releases

Packages

Contributors

Languages