Roadmap · October 2026

What ships today,
and what comes next.

Mneme is the open-source decision and control layer for agentic software development. Architecture is the first decision domain. The enforcement core, the local Decision MCP server, and several coding-agent integrations ship today. The current priority is external evidence: real repositories, design partners, and reproducible architecture-compliance results, ahead of broader expansion.

Install Mneme See the demo Request a pilot
Current roadmap

Current, next, research, and deferred.

Every item on this page carries one of three statuses. Shipped means released and usable today, either in the mneme-hq package or installed from the public Mneme repository. Planned means scoped work that has not shipped. Conceptual means a research direction with no scheduled work and no support claim.

Now Current focus
External validation and adoption

The enforcement mechanism ships. The open question is whether it produces repeatable value on real repositories, so current work favors evidence over integration count.

  • Design partners and real-repository pilots. The evidence sought from pilots is first-attempt architectural compliance, functional completion, unrequested scope expansion, setup and maintenance friction, and continued use on real projects.
  • Architecture Audit as the evidence surface. Shipped. The Architecture Audit classifies each architectural decision it discovers as Protected, Mneme-ready, Requires modelling, or Guidance. Current work uses that assessment as the starting point for pilots.
  • Reproducible architecture-compliance benchmark. The Governance Benchmark v1.1 methodology and harness are public; full results publish after scenario suite validation. Planned: comparisons across coding-agent workflows with frozen fixtures, deterministic scoring, and explicit treatment and control boundaries.
Next Planned
Strengthen the review and ingestion surfaces

Planned work makes the existing enforcement layer easier to apply. None of it creates a parallel policy system, and none of it has shipped.

  • Architecture review workflow. The Mneme Skill ships today in the Claude Code plugin, which loads from the Mneme repository. Planned: evolve it into a fuller executable architecture-review workflow over existing retrieval and deterministic checks. It remains an interface over those checks, not a second policy engine.
  • Confluence ADR ingestion. ADR import from repository files ships today. Planned: ingest decisions that already live in Confluence. Ingestion stays separate from enforcement semantics.
  • Migration-aware Architecture Audit. Planned: extend the Audit for long-running migrations where legacy and target architectures coexist. This is an extension of the Audit, not a separate migration product.
  • Reference review consumer validation. The local Decision MCP server ships today. Planned: validate one external code-review system retrieving applicable decisions through it. This is a compatibility test, not an integration claim, and enforcement stays with Mneme.
  • Deep Agents capability proof of concept. Planned: test whether file mutations and nested agents can be checked reliably. No support claim is made until that is validated.
Research Conceptual
Evidence first

These are research directions, not integration claims and not scheduled work. Each has to show a useful control boundary before it moves up.

  • Shared multi-agent constraints (Slack Code). Whether one project's decisions can apply to multiple coding agents and sessions inside a shared harness.
  • AWS AgentCore benchmark. AgentCore as a controlled execution substrate for architecture-compliance benchmarking, before any deeper integration is considered.
  • Salesforce Skills interoperability experiment. Whether Mneme adds project-specific architecture compliance alongside Salesforce domain Skills. This is an experiment, not an integration claim.
  • Claude Managed Agents revisit. An earlier evaluation produced a partial result, and Mneme does not claim support. It is revisited only if the platform exposes a boundary where a change can be checked and blocked before completion.
  • Decision domains beyond architecture. The Decision Index is designed to support decision domains beyond software architecture, including security, platform, and compliance. No dedicated domain beyond architecture is shipped or scheduled.
Deferred Not scheduled
Waiting for user pull

Broader platform work stays behind external validation. Nothing here is shipped or planned.

  • EventCatalog graph enrichment. Deferred until there is a validated hypothesis for it.
  • Team, organization, and cross-repository policy. Shared policy synchronization and cross-repository aggregation.
  • Generic hosted control plane. A hosted MCP or HTTP service and broad SaaS administration. The released MCP interface is local and stdio only.
  • Higher-level policy DSL. A policy language beyond the current typed rules.
Shipped foundation

What the roadmap builds on.

Today Shipped
Available today, current release mneme-hq 0.10.0
  • Deterministic enforcement. ADR-based decisions, applicability, deterministic rules, agent context, and guide, warn, or block verdicts traced to the source decision. See the decision model.
  • Enforcement integrations. Native adapters for Claude Code, Claude Agent SDK, Google Antigravity, Codex CLI, LangChain / LangGraph, and Kiro. Validated compatibility with Paperclip. OpenCode and Hermes Agent remain experimental. Some adapters, including the Claude Code plugin, install from the Mneme repository alongside the package.
  • Rules export and CI gates. Cursor rules export, GitHub Actions and GitLab CI gates, and ADR import.
  • Architecture Audit and protection activation. The Audit and the mneme protect commands, in the released package since 0.7.0. See protection activation.
  • Persisted Decision Index. Decisions are stored in project memory as immutable versions with a lifecycle (active, superseded, deprecated, inactive); only active decisions are enforced or used by check and test_query, while MCP reads return every decision labelled with its lifecycle. An ADR edit can’t silently drop a protected rule binding. Existing memory moves over with an explicit migration that is lossless or refuses, since 0.10.0. See decision-index migrate.
  • Local Decision MCP server. Six retrieval and proposal tools over local stdio, since 0.9.0, reading the Decision Index since 0.10.0. Proposals are non-authoritative; no MCP tool can accept, reject, or activate a decision.
  • Human decision proposal review. Inspect, accept, and reject proposals through the separate mneme decision authority CLI. See decision proposals.

Support status is evidence-based and changes independently of roadmap priority. The integrations page carries the current public support claim for each tool.

Test Mneme on a real architecture.

Start with the CLI, inspect the demo, or bring a repository into a pilot.

Install Mneme See the demo Request a pilot

Roadmap reflects current priorities as of October 2026 and may change based on pilot evidence and user feedback.