MemoPipe — Automations that learn from your team.
Today, workplace automations are built in isolation. A workflow that works perfectly for one employee may fail for another because their tools, processes, or situations are different. As a result, teams waste time rebuilding automations, troubleshooting failures, and losing valuable knowledge.
MemoPipe turns individual automations into shared, continuously improving team assets.
We built a collaborative platform where employees can discover automations created by their teammates, test them against their own real work situations, and improve them without breaking the original workflow.
Using Screenpipe, MemoPipe analyzes a user's local work history, identifies relevant situations, and recommends existing team automations. Users can then replay these situations in a controlled environment and inspect each execution step.
When an automation fails, an AI agent analyzes the differences between expected and actual results, explains the issue, and proposes a correction. The user can review and test the modified workflow before submitting it to the original creator for approval.
Once approved, the improved automation becomes available to the entire team.
Our prototype demonstrates this process through an IT access-management scenario, including workflow discovery, execution, AI-assisted debugging, human review, and collaborative versioning.
MemoPipe makes automations reusable, adaptable, and more reliable as teams use them.
MemoPipe uses a two-component architecture: a local Screenpipe agent and a collaborative web platform.
Local agent: Built with TypeScript and Bun, it connects to Screenpipe's local API to retrieve permissioned work observations. It identifies relevant workflow situations and sends structured, minimized evidence to the team hub. Screenpipe data remains local until the user authorizes the relevant sharing.
Team hub: Built with Next.js, React, TypeScript, and PostgreSQL (PGlite for local development), it manages automation discovery, scenario testing, execution traces, editable workflow copies, contributions, approvals, and version history.
AI integration: OpenAI-compatible models analyze failed executions, explain discrepancies, and propose structured workflow corrections. Generated changes are validated against a restricted automation DSL and tested before they can be submitted for approval.
Testing environment: A fictional IT service-management application, Nimbus, provides controlled scenarios for demonstrating workflow execution and correction without exposing real employee data.
The key integration is Screenpipe's ability to capture actual work context, enabling MemoPipe to recommend and validate automations against situations encountered by individual users rather than relying exclusively on manually written test cases.
The prototype combines local-first observation, controlled execution, AI-assisted debugging, and human-approved collaborative improvement.