Skip to content
trags
Install Poppy

Where information takes root

Trags builds the software around AI models: what they remember, what they can reach, how they run. We started with memory.

A poppy illustration with branching stems and small leaves on a green background.

Why we started with memory

A coding agent in Cursor, a meeting agent in your calendar, a browser agent doing research: they all lose their context when they leave one session, one tool, or one provider.

Memory is what your agents are missing.

One store your agents write to and read from, whichever model you run and whichever machine you are sitting at.

Poppy is how you get it: a free CLI that captures and recalls on your own machine, offline, with nothing leaving it until you say so. Turn on sync and Trags carries the same memory to every machine you sign in on.

Start with Poppy

Poppy is the free, local-first CLI. Capture and recall from your terminal, and serve the same memory to every MCP agent on your machine.

How Poppy works
$ poppy remember --type preference "prefer pnpm in monorepos"
Remembered (preference): prefer pnpm in monorepos

$ poppy recall "monorepos"
  prefer pnpm in monorepos
    preference | 2026-09-28 | score: 2.47
Say it once, in plain prose.
A memory as Poppy stores it: one row, the same fields every time, whatever you were working on.
id
mem_5a1ed551b7c4
type
decision
content
We pin the http client at >=4.2,<5 because 5.0 removed the retry hook we build on and renamed the timeout option; moving to the 5.x middleware API is tracked as its own piece of work.
project
checkout
source
claude-code
captured
2026-06-11 09:32 UTC
A second flower print, the Poppy mark among a different growth of stems and leaves, printed straight onto the page.

What we study

We study how memory systems for AI agents should be evaluated. Finding the right fact is one test. The question we are working toward is whether memory makes an agent better at its task.

  • One loop for every system. We run every memory system, our own included, through the same test loop, with the same model writing the answers, so the memory system is the only thing that changes.

  • Memory that changes over time. We test whether a system keeps its memory correct as facts change: when a new decision replaces an old one, when the answer depends on the date, and when several memories should become one.

  • The surface agents actually use. We can test a system through its own MCP tools, the same surface a real agent touches, not only through its code library.

None of this work uses your memories. We test on public test sets and on test data we build ourselves. We are writing this work up as a paper, and results will appear here once it is published.

What you can count on

  • Local-first

    Trags memory lives where the agent does. Poppy keeps it in one file on your machine, and sync is opt-in.

  • Never used for training

    On every plan, free or paid, we never use your memories, the embeddings derived from them, or your usage to train or evaluate models.

  • Open source

    Poppy is free and Apache-2.0 licensed. Its source code is public on GitHub.

  • Stored in the EU

    When you sync, your memories and account data are stored in the EU.

Roadmap

What you can use today, and what comes next.

  1. Poppy

    Available

    The free, local-first CLI. Capture and recall from your terminal, and serve the same memory to every MCP agent on your machine. One SQLite file, works offline. Browse and edit that file in a local dashboard with poppy ui.

  2. Trags Cloud

    Available

    Sync and a dashboard. The same memories on every machine you sign in on, with provenance for each one.