Open source · Works with any AI IDE Written in Open Knowledge Format

Your agent doesn’t need more context.
It needs the right context.

Momentum keeps your project’s memory and state as Markdown files in the repo — so every session your agent retrieves the right context, hallucinates less, and turns a vague request into a plan that’s far better grounded: what to change, where, and how.

$ npx @limina-labs/momentum@latest init

init asks which agent to set up — or pass --agent claude-code | codex | opencode | antigravity

No daemon No lock-in Uninstall = delete the files
Same commands, same workflow — every agent
opencodeAGENTS.md
Claude CodeCLAUDE.md
CodexAGENTS.md
AntigravityAGENTS.md
CursorPlanned
Gemini CLIPlanned
The problem

Without the right context, your agent guesses.

Ask an agent to fix a bug and it greps around, fills the gaps with assumptions, and drafts a plan on shaky ground — burning tokens rediscovering what your project already knew. Momentum hands it the project’s real state up front, so it reasons from facts instead of guesses.

Retrieves the right context. From status, history, and decisions already on disk — not whatever a search happened to surface.
Hallucinates less. Fewer confident guesses, fewer wrong turns, fewer tokens spent re-deriving what was already known.
Plans with grounding. A concrete action item — what to change, where, and how — anchored in real project state, not a blank-slate guess.
Without momentum
Greps the codebase, hopes it found the right files Fills the gaps with assumptions — hallucinates Plans on shaky ground, burns tokens rediscovering
With momentum
Reads project state first — oriented in seconds Reasons from recorded decisions, not guesses Produces a well-grounded, actionable plan
13
Autonomous agent rules
15+
Slash commands
0
Background processes
~15min
Zero to a verified release
What changes

Better context in. Better work out.

The state lives with the project, so the agent starts every task already knowing what matters.

Less hallucination

Your agent reasons from recorded facts — status, decisions, history — instead of confident guesses. Fewer wrong turns, fewer retries.

Higher-signal context

The relevant history and prior decisions surface first — not whatever a blind search returned. Less noise in the window, less token waste.

Better-grounded plans

A vague ask becomes a concrete plan — what to change, where in the codebase, how to verify. Not infallible, but anchored in real project state instead of a guess.

Based on hands-on use — momentum reduces guesswork and re-derivation. Your mileage will vary by project and agent.

The phase loop

One repeatable unit of work

Every phase moves through the same five stages. Your agent drives all of it — and stops only at the one gate that matters: the merge and release.

Phase 13 · active
Brainstorm
Scope through dialogue. Disk writes gated.
Plan
Groups, tasks, acceptance criteria.
Execute
Commit per group. Push silently.
Verify
No completion without evidence.
Release
Tag & publish — you approve the merge.
What you get

Open-source primitives that keep state coherent

Across days, branches, and sessions — the building blocks your agent uses to never lose the thread.

Phases

Plan, execute, verify, release. Every phase is a unit of work with its own brainstorm, plan, tasks, and history.

Backlog

Bugs, features, tech debt, and enhancements — with priorities and per-item context the agent files itself.

History

Append-only log of decisions, discoveries, and scope changes. The “why” outlives any single session.

Rules

Thirteen autonomous agent rules — orient first, verify before claim, log every decision. No enforcement overhead.

Skills

Slash commands your agent runs automatically — start, complete, sync, review, validate.

Ecosystem

Coordinate multiple related projects from one agent session — opt-in, additive, zero extra setup by default.

How it works — no daemon, no lock-in

Markdown files your agent already knows how to read

Momentum isn’t a framework or a service. It’s .md files and conventions your AI IDE picks up the moment it opens your repo.

01

init writes files

Your agent’s instruction file, a specs/ folder, rule & command definitions, and git hooks — straight into your repo.

02

Your IDE reads it

Your AI IDE already loads that instruction file every session. Momentum fills it with a structured workflow and a map of where state lives.

03

The agent maintains it

Reads status.md to orient, logs to history.md, tracks work in backlog.md. Automatically.

Every file is Markdown in the Open Knowledge Format (OKF) — so your project’s context stays portable, human-readable, and interoperable across tools instead of locked to one vendor.

Fully compatible
claude-code — my-project phase-13
> /start-phase

Creating branch phase-13-site-polish...
Updating status.md → Current Phase 13

Group 0  Mermaid tooling + baseline      
Group 1  Landing rewrite                 
Group 2  Docs deepening                  

history.md  [DECISION] glass hero over flat
tasks.md    12/12 [x]   push → origin

STOP — merge + release gate. Awaiting your OK.
> _
Watch the discipline land

It narrates. You stay in control.

The agent proceeds silently unless interrupted — but every group commits cleanly, every decision lands in history, and the only hard stop is the merge and publish. Those are yours to approve.

Clean conventional commits, per group
Tasks marked done only with passing evidence
Merge to main never happens without your OK
15+ slash commands

Your agent already knows how to drive this

Phases, backlog, history, doc-sync — all run as slash commands. The three you’ll reach for most:

/brainstorm-phase

Plan the next phase through structured dialogue. Disk writes are physically blocked until you approve.

/start-phase

Begin implementation. Autonomous execution per the contract — stops only at the merge + release gate.

/complete-phase

Verify, finalize, tag, publish. The Rule 12 evidence gets captured into retrospective.md.

The 13 rules

Discipline preserved — without enforcement overhead

Thirteen autonomous behaviors your agent follows in the background. Three of the highest-leverage:

All 13 rules
Rule 6

Git Lifecycle

Feature branch auto-created. Conventional commits. Push silently. Never merge to main without your explicit OK.

Rule 8

Record Phase History

Every meaningful decision, discovery, or scope change lands in history.md the moment it happens. The “why” outlives every session.

Rule 12

Verify Before Claim

No completion without evidence. Run the test, read the output, mark done only if it passes. Closes the “should work now” failure mode.

When one repo isn’t enough

Orchestration across your projects

Single-project is the default — zero extra setup. When related repos need to coordinate, three layers let one agent session work across all of them.

Swarm

Sustained delivery

Drive a whole cross-project feature end-to-end from one session. Conductor + per-project supervisors, wave-ordered, multi-hour. Survives session kill via disk-only resume.

reads + writes

Cross-project actions

One-shot moves

Four single-step verbs the agent runs per task. One move and done — no persistent state, no waves.

scout dispatch handoff continue
reads + writes

Ecosystem

Foundation — shared data

Durable substrate both tiers above read and write: ecosystem.json members + dependency edges, initiatives/ feature tracking, and a daily session log.

Built for the way you work

One discipline, every role

Solo builder

Your project doesn’t forget what it was building when you stepped away for a week. Pick up exactly where you left off, with the agent already oriented.

Tech lead

Bring discipline to agentic AI work across your team. Every decision logged, every phase planned, every release tagged — without enforcement overhead.

PM exploring agentic AI

Read the why behind every decision your engineers’ agents make. Phases, decisions, and history are first-class artifacts — not lost in chat transcripts.

Give your agent a memory
that outlives the session.

One install, and your agent drives the whole loop from files it already knows how to read. Open source, free, no lock-in.

$ npx @limina-labs/momentum@latest init

init asks which agent to set up — or pass --agent

Read the docs Star on GitHub