Systems

Agent DB Experimental

A Markdown knowledge base that lets AI organize itself.

A file-based knowledge-base system: AI agents read and add content, a KeeperAgent sorts, splits, merges and indexes them on its own. Everything is atomic, versioned and fully undoable – works standalone as a plain file collection too.

Request advice How it works

File-based – no server, no database, just Markdown files.

Agent DB: file-based AI knowledge base with KeeperAgent organizing content

What is it?

Agent DB is a file-based Markdown knowledge base that an internal LLM agent – the KeeperAgent – organizes on its own. External AI agents read freely and drop content into an intake folder; the KeeperAgent places, splits, merges, trims and indexes them. Everything is atomic, logged, versioned and undoable. On any inconsistency the database freezes (HALT) rather than acting on a corrupt state.

  • Pure Markdown files – no server, no database, works with any file tooling or Git
  • The KeeperAgent alone decides placement, splitting, merging and indexing
  • Agents read freely (never blocked); writes go through the KeeperAgent
  • Every change is atomic, logged, versioned and undoable

Features

Autonomous organizing

The KeeperAgent places, splits, merges, deduplicates and indexes – driven by a user-authored schema.

Durable links

ID-based references survive every move, merge or split. Alias files ((lnk)) place objects at multiple locations.

Fully undoable

Every content change is versioned; deletions archived. ui_undo reverses any action via the log.

Safe by construction

Atomic writes, global lock, HALT on inconsistency. The DB freezes rather than acting on a corrupt state.

Indexed search

Chained index files with routing summaries. db_search traverses the index via LLM scoring.

How it works

  1. Content is dropped into /_in (or sent directly via db_store)
  2. Dedup check: exact re-drops are detected and skipped
  3. The KeeperAgent places: create, update, merge or refuse
  4. On a full run: restructure the neighbourhood + housekeeping
  5. Flow: Intake → Dedup → Place → Restructure → Trim → Lint → Index

The KeeperAgent makes all organizational decisions – placement, splitting, merging, indexing. External agents read freely and write via db_store; they never move files themselves.

The schema (_schema.md) is user-authored and drives the KeeperAgent: Domain sets the atomic grain, keeperStyle the willingness to act. Split and merge share one grain – merge never undoes a grain-justified split.

What for?

Agent memory

Structured long-term memory for AI agents – they read and write, the DB organizes.

Personal knowledge base

Works without AI as a plain Markdown collection – the KeeperAgent is optional.

Team knowledge

Git-based knowledge base for teams – every change is traceable and undoable.

Documentation & reference

Technical docs that structure themselves – the KeeperAgent keeps them consistent.

Tailor-made for you

Agent DB is modular by design, tuned to your use case via schema and configuration:

  • Domain defines what an atomic object is – drives split and merge behaviour
  • keeperStyle (conservative / balanced / aggressive) controls willingness to restructure
  • deletionPercent controls content trimming (0 disables)
  • naming sets the naming convention (kebab-case, snake_case, Title Case, natural)
  • LLM model and temperature freely chosen – temperature: 0 for reproducible runs
  • Cron schedule or manual trigger – full control over maintenance intervals

Safety & consistency: Every write is atomic (temp → fsync → rename). A global lock serializes writes; reads never block. On inconsistency the DB freezes (HALT) – only diagnose and recovery tools run. Lint detects and auto-repairs structural issues on every run.

See it in action

I’m happy to show you Agent DB on an example setup – or set it up for your use case.

Request a demo

At a glance

TaskAI-organized file-based knowledge base
FormatMarkdown files + resources (images, PDFs, CSVs)
AIKeeperAgent (LLM) – model freely chosen, temperature configurable
Safetyatomic writes, global lock, HALT, versioning, archiving, undo
Actionfile-based (no server) – cron or manual
Interfacesdb_store, db_search, db_read, db_process, db_lint, ui_undo and more

Frequently asked questions

What does it cost?

Setup at a fixed price · plus ongoing usage costs (AI model/APIs). The price depends on scope – the initial consultation is free.

Do I need a server?

No. Agent DB is purely file-based – a folder of Markdown files. Back it up with ordinary file tooling or Git. The KeeperAgent runs on your machine or server.

Does it work without AI?

Yes. The files are plain Markdown – readable and editable with any text editor. The KeeperAgent organizes, but isn't required to use the data.

What happens on errors?

On inconsistency the database freezes (HALT) and blocks all writes. Only diagnose and recovery tools continue. After fixing the cause, HALT is cleared manually.

Can I undo changes?

Yes. Every content change is versioned, every deletion archived. ui_undo reverses actions via the log – with an all-or-none preflight.

Which AI is behind it?

The LLM model is freely chosen (cloud or local). For reproducible runs, temperature: 0 is recommended. The model used is recorded per action.

More than storage: Agent DB is one example of my automation and self-improving AI – systems that organize themselves and learn from their inputs.

Agent DB for your use case

Initial consultation free of charge – on-site in Bamberg/Nuremberg or online.

Get in touch