local-first · runs on your Mac · open-source memory

Nocta watches how you work.
Then does the work.

Nocta records your screen locally and compiles it into memory an agent can trust - so the repetitive work you do every week runs itself, with one click to approve.

Download for Mac Book a demo

open source ★ - stars - installs

watched tuesday · ran friday

What Nocta saw you do

app: Chrome · vendor portal

routine: pull new invoices → rename → file

steps: 7 · same every week

fields: inv no · amount · due date

evidence: frames 99871..100147

What Nocta ran for you

log indone ✓

pull 3 new invoicesdone ✓

rename + file all 3done ✓

anything unusualhalts & asks

same steps as last week · you clicked approve once

How it works

Watch. Compile. Run.

Agents today start every task blind: they know what you typed into them, not what you actually do all day. Nocta closes that gap - no integrations to wire up, no prompts to engineer.

01 · WATCH

It records your screen, locally

Nocta captures which apps, pages, and fields your work moves through - on your machine. Typed content stays out by default, and nothing is uploaded anywhere.

02 · COMPILE

Your day becomes structured memory

Deterministic code - no AI in the loop - compiles the capture into activity frames: a bounded, auditable record of what you did, with evidence behind every line. The same day always compiles to the same memory.

03 · RUN

Repeated work runs itself

The routines you repeat become workflows Nocta replays: same steps, same fields, grounded the same way every time. You approve; it executes. Anything unexpected halts and asks.

No integrations. No prompt engineering.

Your work is the spec

The memory layer

Memory an agent can trust

Most agent memory is written by a model - expensive to build, different every run, and free to invent things that never happened. Nocta's memory is compiled, not imagined: pure code turns your capture into the same structured record every time.

context block · compiled from today0 tokens to build

USER ACTIVITY (2026-07-04 · measured from screen
capture, no interpretation):
coverage: 09:12-20:42 · 342 active min · 11 apps
away: 12:30-13:15 (45m)
- 09:12-09:58 Cursor (46.2m): main.py · api
- 10:01-10:44 Chrome/github.com (41.3m):
  pull_request:acme/api#412 · code:acme/api
- 20:24-20:42 Chrome/linkedin.com (18.0m):
  people_search:cto berlin ×2 · profile:john-doe

86×

smaller than raw capture - fits in any prompt

98.4%

day-recall accuracy, vs 66-80% for AI summaries

0 tokens

to compile a day, and to replay a matched routine

<1s

to rebuild a full day of memory, byte-identical

activity-frames on GitHub Read the research paper →

The memory layer is open source (MIT) and MCP-compatible - any agent can use it. Numbers above are measured on a 51-day real-work corpus; method and harness are in the paper.

Why you can trust it

Calm by design

Local-first, by architecture: Capture, memory, and execution live on your machine. Nothing routes through our servers; you choose what an agent ever sees.
Deterministic memory: Your history is compiled by code, not written by a model - so it can't hallucinate a meeting you never had. Every line carries evidence pointing back to the raw capture.
You approve every run: Workflows replay the steps you were watched doing - and halt the moment something looks different. One click to approve; nothing irreversible happens on its own.
Open where it matters: The recording and memory layer is public code your team can inspect: activity-frames on GitHub, with a research paper describing the method.

Where it earns its keep

Start where the repetition hurts

Nocta shines on the work you do the same way every week - across any app, no API required.

For AP teams

Portal work, done

Invoice collection, statement reconciliation, ERP filing - watched once, run weekly, on your machines. Built for finance data: it can't spend anything, and it halts rather than guess.

See Nocta for AP teams →

And beyond