Typed memory for AI agents.
Projects, rules, and decisions you can return to in the next session.
Start with the PDF guide
The idea, the setup, and everyday use — in one readable document.
Why BURO · How it works · Get started · Documentation
You return to a website project a week later. The agent needs to know where the source lives, which server runs it, what you decided last time, and which changes need your approval. BURO keeps those facts in a record the agent can open when you name the project.
Each work object gets its own memory. Projects, services, devices, books, or research can have their own records. Memory lives independently of project folders and can be shared across devices. The agent loads the records needed for the task and maintains useful facts as the work progresses.
A record has a type, and that type defines the fields, their formats, and what belongs in them. This gives the agent a place to look for a fact and a clear place to save a new one.
For a website project, a record might look like this:
id: my-site
name: My website
kind: project
summary: A personal site with articles and notes.
locations:
- host: laptop
path: /work/my-site
important:
- Ask me before publishing.
avoid:
- Do not put runtime data in the source tree.
read_first:
- README.mdThe bundled types require only id, name, and kind. Start with the facts
you actually need.
What can a record remember?
| Field | What belongs here |
|---|---|
summary |
What the object is and why it exists. |
important |
Confirmed decisions, requirements, and approval conditions. |
avoid |
Things the agent should not do. |
locations |
Source, running installations, URLs, and their devices. |
read_first |
Documents to read before working. |
related |
Connected objects, documents, and useful links. |
notes |
Other durable facts. |
aliases |
Other names you use for the same object. |
Types can also define their own fields. Field guides explain how to use them.
Three types are included:
| Type | For |
|---|---|
project |
Products and independent undertakings: a site, an app, a book. |
service |
Technical components of a working environment. |
host |
Computers and servers. |
Your memory can use your own types. Create book, research, or person
with the fields you need. Definitions live in the database and are edited through
the same reviewed workflow as records. You can start with buro types new book;
the type guide covers the rest.
- Before work — give the agent context. The skill explains how to read and
maintain memory. A session connection runs
buro currentand supplies the current device's record, when present, and a bounded list of associated objects. The supplied System prompt block can keep the core rules in every request. - During work — open the relevant records. When you ask to continue a project, the agent calls BURO through the terminal and reads its facts, decisions, and constraints. It can follow related records as needed.
- After work — keep what will matter later. By default, the skill tells the agent to save useful confirmed decisions, new facts, and corrections. Each change goes through a YAML draft, a diff, and a push. When there is nothing useful to add, there is nothing to write. You can change this policy.
BURO provides the structure, tools, and instructions. The quality of autonomous memory still depends on the agent: choosing what to remember, checking facts, and recognizing when saved knowledge is out of date.
Requires Node.js 24.14 or newer.
npm install -g \
https://github.com/Krablante/buro/releases/download/v3.1.0/buro-3.1.0.tgz
buro init
buro connect opencodezReplace opencodez with opencode, codex, or claude for your harness.
Connect installs the skill and session context connection from the same
package.
After connecting:
- OpenCode / OpenCodez: restart the app to load the plugin.
- Codex: review and trust the SessionStart hook in
/hooks, enable hooks if needed, then start a new session. - Claude Code: start a new session with SessionStart hooks enabled.
For OpenCodez, using both the skill and the System prompt block is recommended: the prompt keeps the core rules present, and the skill explains the workflow. Add the prompt yourself in settings; connect does not edit your System prompt.
You can now ask the agent to use BURO for a project and keep its memory current.
To open a record yourself, run buro my-site; buro help shows the commands
and your memory's active definitions.
A different profile, another agent, or manual editing
Choose a profile explicitly:
buro connect opencodez --path /path/to/profileFor a harness without a supplied connection, add the instruction block to a file:
buro connect file --path /path/to/AGENTS.mdThat file connection adds instructions; automatic session context needs its own setup. See agent connections.
You can also edit memory yourself:
buro draft new my-site project
# Edit the printed YAML file. Keep its __buro metadata.
buro draft diff /tmp/buro-draft-EXAMPLE/change.yaml
buro draft push /tmp/buro-draft-EXAMPLE/change.yamlUse the actual printed path for diff and push. For an existing record, start with
buro draft pull my-site. Failed pushes keep the draft; conflicting changes
require a fresh one. See records and types.
buro update --check
buro updateUpdate installs the latest stable GitHub release on this device, applies compatible bundled type updates, and refreshes existing skills, plugins, and hooks. Custom types are preserved. Restart running agents and the BURO API afterward. Versions before 3.1 need one manual npm upgrade; older custom profiles need connect once to register their paths. See updating BURO.
Local use is a CLI and a SQLite file. For shared memory, run the BURO HTTP server on the database device and point clients to it. Clients keep no database copies. Source code, records, and configuration have separate locations, so updating the package preserves your memory.
Changes are checked for valid fields, formats, references, and conflicting revisions. A successful snapshot is required before every mutation; local init also sets up scheduled backups. BURO runs no model or background chat scanner. You and the agent establish which facts are true.
The shared server is intended for trusted access through a private network, VPN, tunnel, or protected HTTPS proxy. It has no built-in authentication or TLS; storage and backups are unencrypted. See operation and recovery.
The PDF guides at the top walk through the whole system. These references cover individual tasks:
| Read about | English | Русский |
|---|---|---|
| The complete guide in Markdown | Guide | Руководство |
| Types, fields, record edits, v2 migration | Memory model | Модель памяти |
| Skills, prompts, profiles, and hooks | Agent connections | Подключение агентов |
| Updates, shared memory, backups, recovery | Operations | Эксплуатация |
| Architecture, API, development, releases | Development | Разработка |
Open source under MIT. Ideas, questions, and improvements are welcome in issues.

