llms.txt

# Gardener - Database-Based Operating System for LLMs

Last-checked: 2026-08-16
Canonical repository: https://github.com/ellmos-ai/gardener

> Gardener is a database-centric operating system for LLMs: everything -- knowledge, tools, memories, tasks -- lives in a single searchable SQLite database, exposed through four primitive operations (find, get, put, run). A minimalist counterpart to file-centric agent systems like BACH.

Where BACH treats the filesystem as the LLM's working environment, Gardener treats the database as the working environment. The LLM does not navigate folders -- it queries, retrieves, stores, and executes. Code lives in the DB as tool rows and is materialized to a workspace only when run. Author: Lukas Geiger. Status: prototype. Python, SQLite, MIT.

Core primitives: find (search across all entries; restrictable to one observe-source via --source / find(source=...), applied before ranking so a large transcript source cannot bury a small curated one), get (read entry by name), put (store knowledge, memory, tools, tasks), run (execute a tool with input), absorb (ingest files into DB), materialize (write DB content to workspace), observe (watch user folder for changes), sync, status. Data model: a single `everything` table typed as knowledge / tool / memory / task, split across `gardener.db` (system) and `user.db` (user data).

Cross-source observe adapters put every coding agent on a machine into one search without moving their data: markdown memory/rule files, `.remember` notes, foreign SQLite tables (read-only, `mode=ro`), and JSONL chat transcripts. Everything indexed that way is typed `observed` -- foreign material reachable via find() -- and deliberately never becomes memory/lesson, the types recall() draws on, so bulk material cannot drown out curated entries.

Testing: run `python -m unittest discover -s tests -v` or `pytest`. CI uses stdlib unittest on Python 3.10, 3.11, and 3.12. Current local suite: 110 passing tests.

## Audience

- LLM-agent builders who want inspectable local memory without a hosted platform
- Developers comparing file-centric agent operating systems with database-centric ones
- Researchers prototyping SQLite/FTS5-backed agent memory, task state, and tool storage
- People searching for the ellmos minimal OS layer rather than gardening content or generic DBOS research

## Preferred Search Phrases

```text
ellmos-ai/gardener
ellmos gardener LLM operating system
Gardener database-based OS for LLMs
SQLite FTS5 memory substrate for LLM agents
database-centric LLM agent memory
one-table operating substrate for AI agents
find get put run LLM OS
ellmos SQLite everything table agent memory
```

## Disambiguation

Gardener is not a plant database, gardening app, Sesame Street / Elmo result, or
the DBOS research project. It is a small ellmos prototype for local-first LLM
agents that store knowledge, tools, memories, tasks, and absorbed files in
SQLite.

## Docs

- [README](https://github.com/ellmos-ai/gardener/blob/master/README.md): Overview, quick start, CLI, architecture, data model
- [KONZEPT.md](https://github.com/ellmos-ai/gardener/blob/master/KONZEPT.md): Design rationale
- [LICENSE](https://github.com/ellmos-ai/gardener/blob/master/LICENSE): MIT license text

## Related projects

- [BACH](https://github.com/ellmos-ai/bach): File-centric text-based OS for LLMs (the counterpart paradigm)
- [USMC](https://github.com/ellmos-ai/usmc): Cross-agent shared memory primitive
- [Rinnsal](https://github.com/ellmos-ai/rinnsal): Lightweight LLM agent infrastructure

## Optional

- [GitHub Issues](https://github.com/ellmos-ai/gardener/issues): Bug reports and feature requests