All projectsProject 16 of 22 · Build Agents That Remember

DecisionDrift

Your team already made the decision. Decision Drift makes sure nobody forgets it.

The project

Your team already made the decision. Decision Drift makes sure nobody forgets it.

Every day, teams make important decisions in meetings, Slack conversations, and emails. But days later, those decisions are often forgotten, overlooked, or contradicted by someone who wasn't part of the original discussion. This leads to wasted time, miscommunication, and costly mistakes.

Decision Drift is an AI-powered agent built with Screenpipe that acts as a shared memory for your team, proactively preventing actions that contradict previous decisions.

Using Screenpipe to capture the context of what's happening on your computer, Decision Drift automatically extracts important decisions from conversations, stores them alongside their original sources, and continuously compares them with your current work.

When a potential contradiction is detected, the agent alerts you, explains the conflict, provides the original decision as evidence, and suggests a safer alternative.

For example, if a team agrees to keep a legacy API endpoint active until a specific date, Decision Drift can warn a developer who later attempts to remove it in VS Code, helping prevent an unintended breaking change.

Unlike traditional meeting summarizers, Decision Drift doesn't just remember what was said. It turns past decisions into real-time, actionable guidance.

While our current use case focuses on developers, the same approach could help sales, support, and other teams avoid acting on outdated information.

Our vision is simple: stop repeating decisions and start respecting them.

How it’s built

DecisionDrift is a local-first desktop app built with Electron, React, TypeScript and Vite, with local JSON persistence and Zod validation.

Screenpipe provides both the context-capture layer and the AI execution harness. We query its local HTTP API for recent OCR and accessibility text from captured applications, allowing DecisionDrift to understand the employee’s current work without building our own screen recorder.

We also use the Screenpipe CLI to run on-demand AI tasks through a configured MiniMax preset. These tasks extract candidate decisions from meeting notes and compare recent screen context against human-approved decisions. Selected excerpts are processed by the external model through Screenpipe.

DecisionDrift validates the results, checks that supporting quotations match the supplied evidence, and deduplicates alerts within each monitoring session. When a possible contradiction is detected, Electron displays a native notification titled with the relevant decision. Clicking it opens the observed passage, the approved decision and its original source.

🇫🇷 made with love <3 by @nass