Mobius Loop

Mobius Loop is a local-first knowledge loop for AI coding agents working with a Markdown knowledge base and project workspaces.

Why it exists: judgment is the expensive part. Deriving a good decision — with the right context, preferences, and the cost of being wrong and corrected — is costly, and most setups spend it once and lose it: the agent re-derives, the human re-decides, the insight evaporates when the session ends. Mobius Loop is built so that every judgment is captured once and reused many times — so it compounds instead of evaporating. That is the goal the whole loop is organized around. See docs/NORTHSTAR.md for the full thesis (and an honest note on how far real implementations are from it).

The core loop is:

knowledge base -> bridge -> project workspace -> agent work -> writeback -> knowledge base

It is designed for teams that want agents to:

• retrieve relevant prior judgments before starting work, instead of re-deriving;
• carry project context into a local .agent/ workspace;
• capture durable decisions and lessons while work happens, before they evaporate;
• write useful evidence back into a Markdown knowledge base;
• keep automation mechanical while leaving judgment to agents and humans.

Status

Public rewrite draft. The private deployment adapter is intentionally excluded.

This repository should not contain personal notes, private project names, raw session logs, local machine paths, provider keys, or runtime secrets.

Repository Shape

docs/                 Architecture and public protocols
skills/               Agent skills, each with a SKILL.md entrypoint
tools/                Deterministic helpers and checks
fixtures/             Small fake vault/project examples

Public Core Skills

This repository is the Mobius Loop system. Its spine is the loop itself — five skills, with skills/mobius-loop as the controller (which is why the repo and that one skill share a name: the project is named after its core protocol, and that skill is the protocol's entry point). The broader skills ship alongside the loop as a methodology library the loop's agents draw on.

The five loop skills:

• mobius-loop: controller protocol for start/resume/close/writeback flows.
• wiki-retrieval: read-side recall before substantive work.
• wiki-bridge: knowledge-base-to-project context bridge.
• wiki-fy: project/session evidence writeback.
• wiki-fy-shift: scheduled writeback shift protocol.

Around that core, the repository also ships a broader collection of reusable agent-methodology skills — knowledge curation, project-execution governance, spec writing, source ingestion, document and deliverable helpers, and agent-collaboration stances. Each lives under skills/<name>/SKILL.md; see docs/skills.md for the catalog.

Not Included

• private vault data;
• raw conversation/session records;
• personal agent/persona documents;
• node-specific sync profiles;
• private project adapters;
• background service or operating-system automation defaults.

Quick Check

python3 tools/public_readiness_check.py

License

Apache-2.0.