shellbrain 0.1.21

Repo-scoped Shellbrain CLI with explicit evidence-backed writes.

Building a Brain

Shellbrain gives agent sessions repo-scoped long-term memory. It stores what happened, what worked, what failed, and what the human prefers — then retrieves the relevant pieces the moment a similar problem surfaces. Every session compounds into the next.

---

Install

curl -L shellbrain.ai/install | bash

One command on supported machines. The installer provisions the local runtime, installs the Codex, Claude, and Cursor skills, wires the Claude SessionStart hook, and runs shellbrain init for you. On first bootstrap, shellbrain init asks how it should store data. Repos register themselves on first use.

Requirements

• macOS or Linux
• Python 3.11+ required
• shellbrain init asks you to choose one storage mode on first bootstrap
• managed local: Docker installed and the daemon running
• external: PostgreSQL with pgvector
• first init downloads a local embedding model
• Windows is not supported yet

---

Upgrade

shellbrain upgrade

Upgrades the package and reruns init. Skills, hooks, and the managed runtime all refresh in one pass.

The install script also works as an upgrade path — curl -L shellbrain.ai/upgrade | bash if you prefer. Manual alternative: pipx upgrade shellbrain && shellbrain init.

---

Use it

You use shellbrain by launching a skill in your agent.

Codex: Use $shellbrain-session-start to get up to speed in this repo with shellbrain and record durable evidence-backed learnings.

Claude Code: Use Shellbrain Session Start to get up to speed in this repo with shellbrain and record durable evidence-backed learnings.

Cursor: Use shellbrain-session-start to get up to speed in this repo with shellbrain and record durable evidence-backed learnings.

The agent handles everything from there — reading prior context, gathering evidence, writing durable memories at session end. You don't manage any of this directly.

---

Four operations

read retrieves durable memories related to a concrete problem. Re-run whenever the search shifts.

events syncs the active transcript. Returns episode event ids to cite as evidence. Run before every write.

create writes one durable memory. At least one evidence reference required.

update records utility votes, truth-evolution links, explicit associations, or archive state.

The rhythm: read first, events before writes, create/update at session end. Do not rerun shellbrain init every session.

---

Repair

shellbrain admin doctor is the inspect path when something feels wrong.

shellbrain init is the repair path if doctor says repair_needed. The installer already ran it once — you only rerun it to fix things.

If shellbrain isn't found in a tool shell, retry through the shell-specific path the installer configured:

zsh -lc 'source ~/.zprofile >/dev/null 2>&1; shellbrain --help'
bash -lc 'source ~/.bash_profile >/dev/null 2>&1; shellbrain --help'

Fish PATH setup is written to ~/.config/fish/conf.d/shellbrain.fish.

---

DB-backed tests

Live memories and DB-backed tests now use different Postgres hosts.

• managed local Shellbrain keeps durable memories on the machine-owned managed instance
• DB-backed tests and scratch validation should use the dedicated repo-owned test host from docker-compose.test.yml
• scripts/run_tests provisions a disposable test database on that dedicated host by default
• scripts/storage_status shows the live managed target, the dedicated test host, and any legacy local test host that is still hanging around

If you are running managed local Shellbrain, do not leave a stale SHELLBRAIN_DB_DSN export in your shell profile that points at the old local compose database. The machine config wins anyway, and the stale env var just makes the storage layout harder to reason about.

---

Docs

• shellbrain.ai/humans — install, upgrade, getting started
• shellbrain.ai/agents — how agents use shellbrain, with a sitemap to every page
• shellbrain.ai/recall — how the read pipeline works