MomentumMomentumTry the demo
Personal behavior intelligence

From intention to behavior. 

Momentum is a contextual agent, not a chatbot. It turns a goal like “work on my project for 90 minutes” into a structured focus session — then watches your real working context, catches drift the moment it happens, and intervenes before the moment is lost.

No install for this preview — the live demo below simulates a real session end-to-end.

momentum.app / focus session

Focus session

42:18

remaining · 90 min planned

Goal progress

68%

Current context

github.com

Focused

The problem

Every productivity app still assumes you have the discipline it's meant to give you.

Set goals, build tasks, track behavior, read the dashboard, decide what to do next — the tool only helps if you're already disciplined enough to use it correctly. Momentum inverts that: it understands the context you're working in and acts inside it.

Traditional approach

Goal
Application
Advice / statistics
User
Action (maybe)

Momentum

Goal
Context understanding
Agent
Action in the environment
Observed result
Learning

“You should focus more.” — generic advice with no awareness of what you're actually doing right now.

“You planned to finish auth today. You've been on YouTube for 8 minutes.” — then it offers to act.

The agent

Meet the agent — not a chatbot.

No chat window, no prompt box. Under the hood it's a small, boring loop: look at what's happening, decide if it's worth saying anything, and — rarely — ask to act.

Reads context, not conversation.

It never waits for you to type — it watches your goal and your environment.

Silence is the default.

Most sessions get zero messages. It only speaks when a rule actually fires.

Never acts without you.

Every intervention ends in a choice you make, never a command it runs alone.

how the agent actually thinks →

GOAL“finish auth today”CONTEXTtab · idle · elapsedAGENTdecide, then maybe actSTAYS SILENT~70% of sessionsSENDS ONE MESSAGEa specific, not generic, nudgeACTS, WITH APPROVALyou always confirm first
How it works

One loop, closed end to end.

This is the entire MVP surface area: a single loop, demonstrated completely, rather than a dozen half-built features.

01

Goal

“Finish my project MVP in the next 90 minutes.” Momentum parses intent, target date, and daily focus target.

02

Context

The Context Engine turns raw events — tab changes, idle time, elapsed minutes — into a structured state the agent can reason over.

03

Agent

A decision engine evaluates that context against the active session and decides whether an intervention is actually warranted.

04

Intervention

A specific, human message — never a generic nudge — proposing a concrete next action.

05

Action

With your approval, Momentum acts: returns to your workspace, starts a break, or ends the session cleanly.

06

Measurement

Every session produces analytics that feed back into the model of how you actually work.

Live demo — simulated

Run a focus session, right here.

This widget compresses a full session into about 30 seconds. Pick a goal, start the session, then switch into YouTube or Instagram to see Momentum notice and intervene.

new focus session

Work balance

Duration

Work / activity balance

Intervenes sooner — a short grace period before it speaks up.

Today's progress

2 sessions so far

46/ 120 min daily target
Beyond the browser tab

The dashboard is the control center. Two lightweight companions feed it.

The live demo above simulates context in the page. In real use, that same context — and the same intervention loop — is produced by a desktop app and a browser extension talking to this dashboard.

Desktop app

macOS & Windows

Runs quietly in the background to read system-level signals — active window, idle time — the context a browser alone can't see.

Start now

Browser extension

Chrome, Edge & Brave

Watches tab and domain changes in real time and is what powers the distraction detection you just tried above.

Get the extension

Both companions are early access — join now and we'll notify you the moment your platform is ready.

Context engine

Raw events in. A reasoning-ready state out.

The agent never sees a firehose of tab-switch events — it sees a compact, structured snapshot of what's actually happening.

Raw events

USER_STARTED_SESSION
USER_OPENED_SITE
USER_CHANGED_TAB
USER_IDLE
USER_RETURNED
SESSION_ENDED
context.json
{
  "goal": "Build MVP",
  "session_duration": 90,
  "elapsed": 32,
  "current_domain": "youtube.com",
  "productive": false,
  "distraction_duration": 420,
  "last_productive_activity": 180
}
Decision engine

The agent decides whether to speak at all.

Most of the time, the right move is silence. An intervention only fires when the rule actually matches — and the message is generated from the live context, not a template.

the whole pipeline, in one breath →

CONTEXT.JSONwhat's happening nowLLM REASONINGcontext + rule + goalDECISIONspeak? act? stay quiet?MESSAGE OR ACTIONone specific line, not a template

Trigger rule

IF
  focus_session = active
  AND current_domain = distraction
  AND distraction_duration > threshold

THEN
  generate intervention

Generated intervention

“You planned to finish the authentication module today. You've been on YouTube for 8 minutes. Want to head back?”

Agentic actions

Real actions, not just messages.

The MVP demonstrates four concrete actions the agent can take in your environment — each one reversible, each one visible.

Return to work

Hides the distraction and brings back the last productive site.

Open workspace

Opens the project, its docs, and the relevant GitHub repo in one move.

Pause

The user can always ask for a break — the agent never overrides that.

End session

Closes the session cleanly and saves everything for analytics.

Every action above requires explicit approval before it runs. See human-in-the-loop below.
Human-in-the-loop

Momentum should never feel like it's controlling you.

Sensitive actions always require your explicit approval. Every capability the agent has can be scoped, reviewed, and revoked — permissions are a first-class part of the product, not an afterthought.

No action runs silently in the background.

Momentum wants to:

  • Close YouTube
  • Return to GitHub
  • Continue focus session
Session analytics

Every session ends with a real answer, not a streak counter.

Planned vs. actual, where the time actually went, and when you were at your best today.

Today's session

79%

Planned

90 min

Productive

71 min

Distraction

12 min

Idle

7 min

9
10
11
12
13
14
15
16
17

Most productive period: 09:15–10:05 & 11:00–11:45

Main distraction

YouTube

Interruptions

3

Task progress

Authentication → 70%

Next recommended action: finish the OAuth callback tomorrow morning.

Daily reflection

Tracking data becomes behavioral intelligence.

At the end of the day, Momentum turns raw session data into an actual observation and a recommendation — the beginning of a real behavior model, not just a report.

Today

You planned 3 hours of focused work. You completed 2h17.

Your strongest session: 10:00–11:30

Your biggest distraction: YouTube

Observation

Your focus decreased significantly after 16:00.

Recommendation

Schedule difficult tasks before 15:00 tomorrow.

Architecture

Small surface area, real feedback loop.

Six layers, one direction of data flow, and a store that turns every session into training signal for the next one.

User

Sets a goal, starts a session, approves actions

Web app

Next.js dashboard, goals, live session view

Context engine

Events, session state, user goals

AI agent

Reasoning, decision, intervention

Action layer

Browser, timer, notifications

Event / data store

Everything the model learns from

Use cases

Same loop, different intentions.

Goal, context, agent, intervention, action, measurement — the loop doesn't change. What changes is what counts as drift.

Deep work

“I want two uninterrupted hours.” Momentum sets up the workspace, watches context, and intervenes on drift.

Learning

“I want to learn Python.” Momentum clarifies the goal, finds resources, and builds a path with real sessions.

Procrastination

15 minutes into a session, the agent notices you're on social media and names exactly how long you've been gone.

Stuck, not distracted

No tab switch, no progress for 18 minutes. Momentum asks if you're stuck and offers to look at the problem with you.

End-of-day reflection

Completed goals, abandoned ones, distractions, and the periods where you actually did your best work.

Weekly review

Focus +18%, best day Tuesday, 6/8 goals completed, and a pattern: you perform better before 14:00.

Privacy & ethics

This is behavioral data. It has to be treated like it.

The user owns the data. Full stop. That principle is a design constraint, not a footnote.

Transparency

You always know what's collected and why.

Consent

Every capability is explicitly granted, never assumed.

Control

Disable any context source — browser, calendar, anything — at any time.

Data minimization

Only what the loop actually needs, nothing collected on speculation.

Local-first

Raw events can stay on-device; only summarized context reaches the model.

No hidden surveillance

A personal tool for you — never a monitoring layer aimed at you.

Where this goes

The MVP proves one loop. The vision is bigger.

From a focus agent to a system that optimizes the gap between the life you want and the behavior you actually have.

Phase 1now

Focus agent

  • Goals
  • Focus sessions
  • Browser context
  • Distraction detection
  • Interventions
  • Analytics
Phase 2

Personal productivity agent

  • Calendar
  • Tasks
  • Email
  • Slack / Teams
  • Notifications
  • Documents
Phase 3

Personal learning agent

  • Research
  • Learning paths
  • Adaptive learning
  • Spaced repetition
  • Resource recommendations
Phase 4

Life operating system

  • Calendar
  • Tasks
  • Browser
  • Work apps
  • Learning
  • Personal goals — one behavior model
Beyond personal focus

The same loop works inside a company, not just for one person.

Momentum's core mechanic — watch real activity instead of trusting what people self-report — happens to solve one of HR's oldest problems: nobody has real visibility, so promotion, training, and mobility decisions ride on one manager's judgment, once a year.

75%

of employees leave without ever getting a single promotion

ADP Research Institute

$35M/yr

spent on review cycles at a 10,000-person company

CEB / Washington Post

14% → 77%

feel inspired to improve after an annual review vs. after continuous feedback

CEB, SHRM

69%

of companies use no AI at all in their evaluation process

industry survey

Activity optimization

The agent reads real workload, task complexity, and project contribution — not a status update someone typed at 5pm.

Talent management

That activity data feeds reviews, targets training at real skill gaps, and surfaces internal mobility before someone quits to get it elsewhere.

The whole thing lives or dies on trust.

42% of monitored employees are job-hunting within a year, vs. 23% of those who aren't. But 90% accept the exact same data collection when it's tied to a concrete career benefit. Same principle as the personal product: transparency and a real benefit for the person being observed come first — never bolted on afterward as a compliance checkbox.

IssueTodayWith the agent
Talent detectionOne manager's yearly judgment callContinuous, objective tracking
TrainingGeneric, based on stated wishesTargeted, based on real activity
Internal mobilityBarely visible, decisions come lateDetected proactively, before someone leaves
Avoidable turnover costLost talent that was never promoted (75% of exits)Reduced through real visibility
Internal mobility also costs 18–20% less than external hiring, and reaches full performance 2–3 years faster — the same visibility gap Momentum closes for one person closes it for an entire org.

Stop measuring intentions.
Start acting on them.

Momentum from intention to behavior — one focus session at a time.