Screenpipe
Records what's on your screen and what you hear, keeps it on your machine, and lets AI agents search it.
by Screenpipe (Negentropy Labs, Inc.) · Built this? Edit this page on GitHub
In short
Screenpipe runs in the background on your computer. When something changes on screen (an app switch, a click, a pause in typing), it saves the screen's text, mostly from the accessibility tree with OCR as a fallback. It also transcribes system and microphone audio and marks who spoke. Everything goes into a SQLite database on your machine. An AI agent reaches it through a local REST API or an MCP server: it can search by keywords, app, window, URL and time, ask for a summary of a time range, or look at one moment in full. Agents can also save short "memories" (facts, preferences, decisions) with tags and an importance score. Scheduled AI tasks, called pipes, can read the history and write notes, skills or workflow write-ups from it.
Who it’s for
People who want their AI tools to know what they saw, read and heard on their computer, including meetings. Also developers and teams building agents or automations on top of a local activity history.
What it remembers
Screen text with its app, window and URL, screenshots, audio transcripts with speakers, keyboard input and app switches, meetings, and memories that an agent or you choose to save.
- Memory lives
- Your machine (sync optional)
- API keys
- None for capture and search; an AI model (local, such as Ollama, or cloud) for chat and scheduled tasks
- Works with
- Claude Code, Claude Desktop, Codex, Cursor, Gemini CLI, ChatGPT (via MCP), Any MCP client
- License
- Partly open sourceScreenpipe Commercial License (source-available, not open source)
- Price
- Free plan; paid Basic and Business plans (the sources don't give rates); Enterprise priced per deployment
Under the hood
- Who writes memories
- The recorder saves screen text and audio transcripts on its own whenever something changes. Memories are saved by an agent (the update-memory tool or REST API), a scheduled task, or you.
- Where they're kept
- A SQLite database on your machine with full-text indexes, plus video, audio and image files on disk. Optional encrypted sync between devices.
- How they're found
- By keywords, with filters for app, window, URL, speaker, time, tags and importance. An activity summary covers a time range, and a timeline shows any single moment.
- Extra AI calls
- None for capture and search (local OCR and Whisper by default). Chat and scheduled tasks call the AI model you choose. Cloud transcription is optional.
How it works
- 1You workScreen, audio and input are captured when something changes.
- 2StoreText, transcripts and media go into SQLite and files on your machine.
- 3Agent asksThrough MCP or the REST API, by keywords, app and time.
- 4Save and learnAgents and scheduled tasks save memories, skills and workflows.
What stands out
- It captures work outside the AI chat: what was on screen in any app, and what was said in meetings, with speakers.
- One local history that any agent can use: a SQLite file, a REST API on localhost and an MCP server.
- Each scheduled task can be limited to certain apps, windows, content types and hours. The README says the limits are enforced in code, not by asking the AI.
- The skill-learning task saves a reusable method only after it sees the same method or correction on two separate occasions.
What it doesn’t do
- Captured history is kept as it was recorded. Nothing fades, strengthens or merges it based on use. Retention is off by default; when you turn it on, it deletes data older than a set number of days (by default only the video, audio and image files).
- Saved memories don't come back on their own, except the in-app agent's short profile of you, which starts every chat. Editing a memory overwrites it with no history, and nothing checks a new memory against older ones.
- It isn't open source. The Screenpipe Commercial License allows free personal, non-commercial, nonprofit, educational and research use, plus a 7-day evaluation. Any commercial use needs a paid Enterprise plan. Older versions released under MIT stay MIT.
- The free desktop app reads back only the last 24 hours of history, in search, the timeline and MCP alike, and allows two tasks of your own. Paid plans lift both.
- It records your screen, audio and keyboard input. Capture and search stay on your device by default, but cloud AI, cloud transcription (Deepgram), sync and integrations send data off it. Product analytics (PostHog) and crash reports (Sentry) are on by default and can be turned off in Settings.
- It doesn't write memories while it records. Saved memories come from an agent, a scheduled task or you.
Measured against human memory
0 built, 9 partly, 2 not of the Field Guide’s 11 mechanisms
Encoding
Salience
Decides what's worth keeping, and how much, as a memory is written.
The in-app agent is told to save stable facts about you and skip task progress, secrets and facts likely to be stale within a week, and a filter rejects secret- or instruction-like text. Importance is a default or what the writer sets. Capture keeps everything except what you exclude. source
Emotion
Gives memories written with feeling more weight.
No emotional weighting. Nothing in the code reads or records feeling.
Storage
Forgetting
Lets unused memories fade over time on its own.
An opt-in retention setting (off by default, 14 days if turned on) deletes captured data by age, by default only media files. Memories don't fade. source
Interference
When a new memory closely matches an old one, notices and lets the newer one win.
The in-app agent is told to update a matching profile fact rather than add a copy. That's the AI's judgement; no step compares a new memory with old ones. source
Retrieval
Retrieval
Remembering something makes that memory stronger.
Being found doesn't strengthen anything. Memories have no use counter, and results are sorted by date or importance.
Association
Links memories, so recalling one brings its neighbours.
Tags the caller picks (such as person:ada) tie memories to screen moments and audio, and a search option lists the tags that appear alongside them. Recall doesn't follow these links on its own. source
Prospective
Brings a memory back at the right time without being asked.
Tasks run on a schedule or when a meeting ends, and the meeting summary is on by default. That runs a task at a time; a saved memory doesn't come back on its own. source
Transformation
Consolidation
Reworks memories between sessions: strengthening, merging, reorganising.
If you turn on the workflow tasks (off by default), an AI pass checks saved workflows against new activity and corrects or enriches them. Saved memories and raw history are never reworked. source
Reconsolidation
Corrects a memory when it's contradicted, and keeps what it used to say.
You or an agent can edit a memory to correct it, but the edit overwrites it and the old wording isn't kept. Sync between devices keeps whichever edit came last. source
Gist
Turns many specific moments into general knowledge.
A skill-learning task, off by default, turns a method seen on two separate occasions into a reusable skill once you switch it on. It learns methods only, not general facts. source
Schemas
Forms beliefs about people and projects, and weighs new facts against them.
The in-app agent keeps a profile of you, up to 20 facts ranked by importance, and every new chat starts with it. It's about you only, not other people or projects, and new facts aren't weighed against it. source
Not every system is trying to be a human memory, so a hollow dot isn’t a flaw. It just says what the system doesn’t do. What the mechanisms are.
Screenpipe’s makers haven’t shared what they’re building next, so only today shows here. Makers can add their roadmap the same way they correct an entry.
Compare Screenpipe with the others, side by sideQuestions
- Who is Screenpipe for?
People who want their AI tools to know what they saw, read and heard on their computer, including meetings. Also developers and teams building agents or automations on top of a local activity history.
- Where does Screenpipe keep its memories?
Your machine (sync optional). A SQLite database on your machine with full-text indexes, plus video, audio and image files on disk. Optional encrypted sync between devices.
- Does Screenpipe forget?
Forgetting: a little. An opt-in retention setting (off by default, 14 days if turned on) deletes captured data by age, by default only media files. Memories don't fade. source
- Does Screenpipe correct a memory when it turns out to be wrong?
Reconsolidation: a little. You or an agent can edit a memory to correct it, but the edit overwrites it and the old wording isn't kept. Sync between devices keeps whichever edit came last. source
- Is Screenpipe open source?
Partly open source. License: Screenpipe Commercial License (source-available, not open source). Code: github.com/screenpipe/screenpipe.
- Does Screenpipe work with Claude Code?
Yes. Works with: Claude Code, Claude Desktop, Codex, Cursor, Gemini CLI, ChatGPT (via MCP), Any MCP client.
- Does Screenpipe need an API key?
API keys: None for capture and search; an AI model (local, such as Ollama, or cloud) for chat and scheduled tasks.
- How much does Screenpipe cost?
Price: Free plan; paid Basic and Business plans (the sources don't give rates); Enterprise priced per deployment.