External Brain: Why We Gave the Assistant Tools Instead of a Giant Prompt

External Brain: Why We Gave the Assistant Tools Instead of a Giant Prompt

When we first built External Brain (https://brain.autobahn.bot), our goal was deceptively simple: stop repeating our project preferences, code styling rules, and architectural decisions to our AI assistant every single morning.

In our earliest experimental prototypes, we did what almost every AI developer does: we created prompt templates. Before each session, a script concatenated the user's scratchpad notes, recent Git commits, and coding rules into one giant preamble.

Within two months, that approach hit an impassable brick wall:

• Context Window Bloat: The preamble ate a large share of the context window before the developer even typed their first message.
• Instruction Dilution: When an assistant receives 4,000 words of background rules all at once, it experiences attention drift. Critical guidelines get overlooked in favor of whatever rule appeared nearest the bottom.
• Stale Assumptions: Monolithic prompts cannot dynamically query or verify whether a rule applies to the specific file currently open.

To solve this fundamentally, we built External Brain as a tool server that any coding assistant can connect to through the open standard those assistants share—and published it as open source.

Here is why that architectural shift changed everything.

From Static Ingestion to On-Demand Tool Contracts

That open standard gives an assistant hands instead of ears. Instead of flooding the model with unsolicited background context upfront, External Brain equips the assistant with a catalog of tools it can call.

External Brain breaks monolithic memory down into 14 purpose-built tools. The four that matter most:

1. 1`brain_start_session`: Opens a task; the task description doubles as the search, and the answer already carries the rules learned before that match it.
2. 2`brain_report_session_outcome`: Closes the task with up to five distilled lessons; External Brain refines them into durable rules and leaves transient noise behind.
3. 3`brain_teach_knowledge`: Records a rule or decision the moment you state it; a new decision can name the one it replaces, which is then retired, never deleted. Duplicates are merged nightly.
4. 4`brain_ask_oracle`: Answers "why did we…?" from the rules and sessions it has, with citations.

The Power of Tacit Memory

What makes External Brain unique is that it does not merely store raw chat transcripts. It extracts tacit knowledge—the unspoken instincts and working principles of your team:

• "We always use static exports for Cloudflare Pages; never introduce Node.js server middleware."
• "In Thai copy, maintain a warm, conversational register without corporate jargon."
• "Ensure all interactive elements have explicit minimum 44px touch targets for accessibility."

When an engineer starts working on a new component, the assistant opens the task with brain_start_session behind the scenes, and only the rules that match that task come back with the answer.

Fully Open Source and Self-Hostable

External Brain is not a proprietary closed-door API. It is completely open-source under the MIT license at https://github.com/bejranonda/ExternalBrain.

You can spin it up locally in Docker or deploy it to your own private server. It works with any assistant that speaks the same open standard—External Brain ships ready-made setups for seven of them, from IDEs to terminal assistants.

Your memory stays on your infrastructure, under your control, learning and sharpening alongside your team every single day.