SiftText

Understanding evolves from evidence.

curl -fsSL https://sifttext.com/install | sh
GitHub

Build an understanding that people and agents can keep developing.

Reconsidered decision 01 / 04

  1. Which supplier can meet our delivery deadlines?

    Both companies have sent quotes and samples. We’ll compare them to choose a supplier, then track whether orders arrive on time.

    1. Use Company A for every order

      Both suppliers’ samples passed inspection, so we chose A at $8.20/unit over B at $10.

      That choice assumed A’s five-day delivery estimate was reliable. The later ruled out A for urgent orders.

      1. Review Company A’s deliveries

        Three of five orders arrived late. Two delayed customer shipments.

        The lower price no longer justifies relying on A for urgent work.

    2. Trial Company B for urgent orders

      exposed missed customer deadlines. B quotes three days, but its delivery performance remains untested.

      Trial 100 units with B. Keep A for flexible orders.

      1. Check Company B’s trial results

        No trial results yet.

        Check delivery dates and quality before giving B more orders.

  1. Why did signups drop in March?

    Weekly signups fell 30% starting March 4. We’re checking each change from that week.

    1. The new pricing page

      It launched March 11, a week after the drop began, so it can’t explain the start.

    2. Verification emails

      A DNS change on March 4 sent verification emails to spam. Fixing it recovered most of the lost signups.

    3. Seasonal demand

      Signups are still 8% below February. Compare with last March before ruling this out.

  1. Why do the nightly tests fail?

    About one night in four the run fails, but a rerun usually passes.

    1. Find the failing tests

      Every failure comes from a test that writes to the shared database.

      1. Run them one at a time

        Each passes alone. They fail when two tests change the same records at once.

        1. Give each test its own database

          Separate databases keep tests from changing each other’s records. Try it on tonight’s run.

  1. Choose a database for the new app

    Its main reports join data from several tables.

    1. Use Postgres

      It runs those reports directly, and the team already operates it.

      1. Load-test with real data

        Report queries stayed under 40 ms at peak traffic.

    2. Use a document store

      It’s quicker to set up, but each report would need data combined in code.

Plug-in memory for any agent.

Your agents keep the same tree when you change models or apps.

Your project’s treeMarkdown + Git
  • Coding agent
  • Chat assistant
  • Research agent
  • Next year’s model
Local
Agents with file and terminal access read and update Markdown files in Git.
MCP
With Plus, any MCP client connects to the same synced tree.

Plans

Core

Free

  • Open source
  • Markdown files with Git history
  • For agents with file and terminal access
View on GitHub

Plus

$20/month

  • Everything in Core
  • MCP access for your agents
  • Cloud sync
Choose Plus

Pro

$50/month

  • Everything in Plus
  • A full interactive browser your agent uses through MCP
  • Writing checks that catch AI-sounding prose
  • Automatic tree organization checks
Choose Pro

Running agents across a team?

Bring your team’s work into one place for people and agents.

Book a demo