AI Brain

The challenge

When you run many tasks through AI agents, every task is a cold start. The agent doesn't know your preferences, your team, the stack on a given project, or the decisions made last week - so you re-explain context over and over. That overhead quietly eats the speed advantage AI is supposed to give.

What we built

AI Brain is a global context brain for tasks: it distills the team's conversations into a structured, queryable knowledge base that any AI agent can read before starting work. Instead of repeating context in every prompt, an agent asks the brain once and gets clean, deduplicated knowledge.

The pipeline:

1. Ingest - pull conversation history incrementally (cursor-tracked, so reruns are fast and idempotent)
2. Distill - Claude filters noise at ingestion and extracts only long-lived facts, merging them into structured Markdown notes
3. Synthesize - automatic linking between people, projects, and topics, plus communication-style analysis
4. Serve - an MCP server exposes tools like get_profile(), get_person(), and search_brain() that other projects and agents call directly

The role of AI

The key design choice is distillation at ingestion, not at retrieval. Claude decides what is worth remembering as data flows in, so the brain stays focused - no junk in, no junk out - and scales cleanly. The knowledge base is plain Markdown (Obsidian-compatible), so it's transparent, auditable, and editable by hand.

For our AI-first workflow this is core infrastructure: an agent picking up a task on, say, a client integration can fetch the project's team, stack, and open questions in seconds - turning every cold start into a warm handoff.

Engineering notes

• Incremental by default. Ingestion is cursor-tracked, so a rerun costs almost nothing and the pipeline can run often instead of as a nightly batch. A brain that is a day stale is a brain agents stop trusting.
• Plain Markdown as the storage format is a deliberate constraint. A vector store would have been easier to write to and impossible to read: when an agent acts on a wrong fact, someone has to open the file, see the sentence and delete it. Obsidian compatibility means that person needs no tooling from us.
• Deduplication is the real work. The same fact arrives in a dozen conversations in a dozen phrasings. Merging those into one note, and correctly replacing a fact when it changes, is harder than extracting it in the first place.
• MCP rather than a bespoke API. Exposing the brain over MCP means any agent in any project can call it without an integration, which is why it became shared infrastructure rather than one tool's feature.

The result

Context setup that used to cost minutes per task drops to a single query, and agents act with the team's real knowledge instead of generic guesses.

Stack: Python, Telethon, Claude (Anthropic SDK), MCP / FastMCP, Markdown / Obsidian.

Service: custom AI agent development