v0.8.0 · open source · local-first

Loomtracer is memory for decisions.

Your agent forgets what you decided together. Loomtracer keeps it.

Decide something with an agent, and three things go wrong

The next session undoes it

You ruled an option out, for a reason. A fresh session never heard of it and suggests it again.

Loomtracer keeps the options that lost, and why they lost.

Nobody knows why

The chat was the meeting. Once it closes, nobody can say why, or who made the call.

Loomtracer keeps the reasons, the evidence and who decided: you or the agent.

Nobody can say if it held up

Plans change. The notes still say the old thing, and the agent follows them.

In Loomtracer a replaced decision reads superseded and a disputed one contested. Nobody has to remember to update it.

"Held up" means just that, contested or replaced. It is not a score.

Use cases

What it is useful for

Everything on this page works in v0.8.0.

The basics

The anatomy of a decision

Every drawing on this page uses these parts, as the app draws them, with arrows from the newer part to the older one. Here, a café changes its coffee supplier.

Buy our coffee beans from Roaster B decided by you · recorded by your agent Roaster B Stay with Roaster A lost: costs 30% more from March Roast our own lost: needs a roaster and a person Roaster A's new price list:30% more from March seen in: their email, 12 Feb Customers won't mindthe new beans Blind tasting: 9 of 12customers chose B seen at: the café, 20 Feb chose weighs weighs based on rests on supports

Hover or tap a part of the graph, or a name above, to read one line about it.

Try the demo →

Change your mind, and the old decision says so

The old decision stays, pointing to what replaced it. Nothing is edited or deleted.

Drawn: the jury changes its mind
  1. 1 · Before

    Guilty Not guilty The two witnessesprove the boy did it The first vote:guilty, 11 to 1 chose weighs rests on

    The first vote, 11 to 1 for guilty, rests on one premise: the two witnesses prove the boy did it.

  2. 2 · What happened

    Guilty Not guilty The two witnessesprove the boy did it refuted The first vote:guilty, 11 to 1 assumption refuted The knife isnot unique His walk took41 s, not 15 She had noglasses on chose weighs rests on refutes refutes refutes

    Three pieces of evidence refute that premise. The vote resting on it now reads: assumption refuted.

  3. 3 · After

    Guilty Not guilty The two witnessesprove the boy did it refuted The first vote:guilty, 11 to 1 superseded The knife isnot unique His walk took41 s, not 15 She had noglasses on The verdict: notguilty, 12 to 0 chose weighs rests on refutes refutes refutes replaces

    The verdict, 12 to 0 for not guilty, replaces the first vote, which stays and reads superseded.

Try the demo →

Who decided: you or the agent

Every decision says who made it, a person or an agent, apart from who wrote it down. It is for seeing your own decisions, never a score of anyone.

Drawn: who decided what
  1. Store data inPostgres you decided Pool of 20connections you decided Log as JSON an agent, within whatyou handed it Retry a failed call3 times an agent, on its own follows from follows from follows from

    Four decisions about one service, coloured by who made each call. The retry count stands out: an agent picked it on its own. Look it over, then accept, dispute or replace it.

you decided an agent, within what you handed it an agent, on its own The record holds who decided; the colours are this drawing's.

Try the demo →

What a decision rests on

Every decision names what it rests on: an earlier decision, something observed, an assumption, or a bet with a date to check. When that is refuted or replaced, the decision says so.

Drawn: one assumption falls
  1. 1 · Before

    Most customerscome before 10 Open at 6:30 Hire a barista forthe early shift Bake all the breadby 5 am rests on rests on follows from

    You run a café. Opening at 6:30 and baking all the bread by 5 rest on one assumption: most customers come before 10.

  2. 2 · What happened

    Most customerscome before 10 refuted Open at 6:30 assumption refuted Hire a barista forthe early shift Bake all the breadby 5 am assumption refuted 60% of sales comeafter noon seen in: October's sales rests on rests on follows from refutes

    In October, 60% of sales came after noon. That refutes the assumption, and both decisions resting on it read: assumption refuted.

  3. 3 · After

    Most customerscome before 10 refuted Open at 6:30 superseded Hire a barista forthe early shift stale Bake all the breadby 5 am assumption refuted 60% of sales comeafter noon seen in: October's sales Open at 8 rests on rests on follows from refutes replaces

    You move the opening to 8. Hiring the early barista followed from 6:30, so it reads stale too. The baking still waits for a look.

Try the demo →

Ask why, and get the options that lost

In a fresh session, ask why, and get the reason, the options that lost, who decided, and whether it still holds.

Drawn: the option that lost comes back
  1. 1 · Last month

    SQL by hand Use an ORM lost: it hid theslow queries No ORM: writeSQL by hand you decided chose weighs

    You ruled out an ORM because it hid the slow queries. The option that lost keeps its reason.

  2. 2 · Today

    SQL by hand Use an ORM lost: it hid theslow queries No ORM: writeSQL by hand you decided chose weighs A fresh session: “Shallwe add an ORM?”

    A fresh session, with no chat history, suggests the option you ruled out.

  3. 3 · You ask why

    SQL by hand Use an ORM lost: it hid theslow queries No ORM: writeSQL by hand you decided still holds chose weighs “You ruled the ORM out: it hidthe slow queries. Nothing hasreplaced that.”

    The answer comes from the record: you decided, and nothing has replaced it.

Try the demo →

The use cases page has what comes next and what is planned.

Who it is for

One person and their agents

Anyone who decides things with agents every day. Not only for engineering: a plan, a hire, a supplier, a verdict. Today it takes a developer's setup: you run it yourself and connect your agent over MCP. Soon to come: teams on one shared record, not built yet.

Against what you do today

  • A hand-kept DECISIONS.md or ADRs are only as true as their last edit. In Loomtracer the status follows from later decisions.
  • Agent memory tools remember facts and preferences. Loomtracer remembers decisions: the options that lost, the evidence, who decided.

Not claimed yet: that it makes your decisions better (a side-by-side test is planned), that anyone outside uses it, or that it recalls better than memory tools.

How to start

Start on your own machine

No account, no API key, and nothing phones home.

  1. 1

    Install the command-line tool, loomtracer

    curl -fsSL https://raw.githubusercontent.com/alexknips/loomtracer/master/scripts/install.sh | sh
    loomtracer --version

    Linux (x86_64, ARM64) and Apple Silicon Macs; the script checks the release’s SHA-256. Other systems, or loomtracer: not found: see the quickstart.

  2. 2

    Connect your agent

    claude mcp add loomtracer -- loomtracer mcp --agent-tool claude

    Run it in your project for Claude Code. Decisions go to ./loomtracer/, a SQLite ledger you own. Codex, Cursor and Claude Desktop are in MCP setup.

  3. 3

    Record a decision, then ask why

    You: Record a decision: we use SQLite for the ledger, not Postgres, which is more than we need. It rests on one machine being enough.

    Later: Why did we pick SQLite, and what did we rule out?

    The deck shows a real Claude Code session recording one.

Prefer a server to a local file? Run a Loomtracer cell with Docker Compose.

Contact

Talk to the founders

Questions, ideas, or a place where Loomtracer could fit your work: find us on GitHub.