Gardener

Landing Page
Main Flow
User onboarding flow
Example usage
Example usage 2
Main Dashboard

Inspiration

An assistant whose memory takes care of itself.

Everyone who uses AI with memory has hit the moment it remembers wrong: the diet you quit, the job you left, the gym you canceled. The longer you use today's AI, the more wrong it gets about you — and you can't see or fix what it believes.

Gardener is the opposite, built from three pieces that turn out to be one product:

One agent you chat with, acting for you on the open web.

Standing watches as visible, steerable conversations. Ask it to "watch Zillow for houses in my neighborhood" and the watch appears as its own chat you can open and steer mid-task — "only 3+ bedrooms, 1500+ sqft." Watches keep running unattended.

Memory that gardens itself. Your steering distills into a plain-markdown preference vault you can read — every fact with provenance. A background lint agent audits the vault against the event log, catches contradictions and staleness, fixes them with receipts (auto-applying high-confidence diffs, queueing risky ones for one-click approval), and publishes its correction changelog publicly with sources.

The more you use it, the better it knows you — the opposite of every assistant today.

Claude Code is super powerful but needs a lot of setup, I wanted to build an agent that has context of its user, can perform actions on their behalf and is almost like a partner or friend. Also I wanted it to work out of the box. I believe every person should have a custom agent harness where the skills, tools, agents and most importantly memory is autolinted. The user shouldn't explicitly give the agent context, it should learn from interactions, thats where Clickhouse comes in.

What it does

The user tells gardener what they are working on and gardener figures out what tools, skills, agents and context to grow for itself based on the user and the context they give the agent. As the user interacts, the garden grows and gardener learns. The biggest feature is AutoLinting where context is both preserved and forgotten (super important) based on the user's inputs and actions (using Clickhouse to track the user's interactions).

How I built it

Claude Code as the main assistant.

Pioneer where I have a fine tuned NER model for inference, synth data based on 577 potential user interactions, The model is focused on finding entities based on the user's chat request. A claude agent is the fallback.

Composio gives me integration with discord to alert the user, calendar integration and gmail.

Render for all things hosting from env variables, a web service, backend and frontend hosting, a blueprint for both the api and limiting and the new workflow function to run hosted loop functions.

Clickhouse to track user actions.

Challenges we ran into

Managing context, preservation of the user's preferences, and I ran out of credits to ship parts of the project but I'm proud that its something I'd actually use!