AI Second Brain

Koda's personal context layer. Where KodaMemory remembers your project, the Second Brain remembers you — who you are, what you're building, your goals, how you like to work, and how you actually write. It's seeded by a short guided interview, keeps learning ambiently as you work, and briefs every agent — in every workspace — with the facts relevant to each task. The result: outputs tailored to you instead of generic.

Overview

Every agent session starts knowing nothing about you. You end up re-explaining the same things — what you're building, who it's for, that you want answers short and direct — over and over, in every pane, in every project. The Second Brain fixes that: it's a per-user knowledge graph of identity facts, and Koda injects the relevant slice of it into your agents' context automatically.

Setting it up — the interview

Open the "👤 You" tab in the side panel and choose Set up my Second Brain. Koda launches a guided interview in an agent pane: a short, conversational intake that covers who you are, what you've built and are building, what you're aiming at, how you like to work, and how you want AI to respond to you — length, tone, format. Answer honestly and in your own words; as you answer, each response is distilled into typed facts and saved to your brain.

• Resumable — pause anytime and pick it back up.
• Re-runnable — run it again later to top up the brain; re-asserted facts merge rather than duplicate.
• Your voice is part of it — the interview explicitly captures how you want responses written, so agents mirror your style from day one.

What it stores

The brain uses six identity-focused node types — the surface that actually changes an agent's output:

How it learns as you work

After the interview seeds it, the brain keeps itself current with ambient learning — no manual upkeep. When an agent session ends, Koda runs a lightweight background pass over what was said and proposes candidate facts about you: a new goal you mentioned, a preference that shifted, a project that changed shape.

The confidence gate

Not everything an AI thinks it noticed about you should be remembered. Candidates are split by how they were learned:

• You said it outright, clearly → saved silently. ("I'm switching the landing page to annual pricing first" → saved.)
• Koda merely inferred it → held in a review digest — never injected, never trusted, until you approve it.

The digest appears in the "👤 You" panel as "✨ New knowledge to review", each candidate with a one-click Keep or Dismiss. It's passive — it never interrupts your work — and dismissed candidates aren't re-proposed. Facts learned ambiently carry a small ✨ auto tag in the viewer so you can always tell what you told Koda versus what it picked up.

Style synthesis

Separately, Koda periodically studies how you actually write — sampling your real prompts and distilling verbosity, tone, formatting and vocabulary into a single, always-current style profile. Agents use it to answer in your voice rather than a default one. It updates in place (the old profile is superseded, not duplicated) and is visible and forgettable like any other fact.

Kept deliberately cheap

• Trivial sessions (a few short messages) are skipped entirely.
• Background passes are capped per hour, and use the same agent CLI auth you already have — no separate API key.
• It's a toggle: "Learn about me as I work" in the "👤 You" panel. Turn it off and the brain only changes when you tell it to.

Reading Queue — feed it links

The other way knowledge gets in is the 📚 Reading Queue in the "👤 You" panel. Paste a link you never had time to read — an X post or thread, or any article — and Koda fetches it in the background and runs a headless distiller over the content. You get three things back, per item, without reading a word yourself:

• Durable learnings — the handful of things actually worth keeping, saved as learning facts you can Keep or forget individually.
• Content angles — ready-to-edit post drafts (a hook + body) built from what the piece said, so an interesting read becomes something you can ship.
• Idea sparks — project ideas the piece suggests, tagged to the project they apply to.

How agents use it

On every prompt you send in an agent pane, Koda ranks your brain against what you're asking and injects only the facts that are relevant to that task — a handful, not the whole file. Ask something unrelated to who you are, and nothing is injected at all; your context stays clean and your usage isn't spent on facts the task doesn't need.

What Koda knows about you (personal context from your AI Second Brain):
  - solo founder, building Koda (an ADE) [person]
  - ship v1 to first paying users this quarter [goal]
  - direct, concise answers — no corporate fluff [preference]
  - writes casual, short sentences, hates buzzwords [preference · style]

One exception is deliberate: ask an agent a question about yourself — "what do you know about me?", "sum me up" — and Koda injects the whole brain, so the answer is complete rather than a top-slice. Agents are also explicitly instructed to treat these facts as the only facts about you and never to invent details beyond them.

Injection is push — Koda decides what's relevant up front. Agents can also pull: the koda_brain_query tool lets an agent search your brain mid-task, the moment it realises knowing who you are (or what you've read) would change its output. It's read-only, Pro-gated, and searches both your identity facts and your Reading Queue learnings — so an agent can draft in your voice or ideate from something you saved without you re-explaining it.

Managing your brain

The "👤 You" panel is the single place to see and control everything the brain holds:

Privacy & tiers

This is the most personal data Koda touches, and the stance matches:

• Local only. The brain lives in ~/.koda/brain on your machine — outside every repo, so it can never be committed or pushed. There is no Koda backend and nothing is uploaded, synced, or telemetered.
• Your own auth. Ambient learning runs through the same agent CLI you already use, under your own login — Koda never proxies a credential.
• Fully inspectable. Every fact is visible in the "You" panel with its provenance; forget is immediate and hard.

Access gating fails closed, like the rest of Koda:

If your subscription lapses, injection and learning pause immediately — but the brain itself stays on your disk, intact, and reactivates when you're back on Pro.

Troubleshooting