Episodes
The raw transcript trail. What actually happened, in order. This is the ground truth everything else is distilled from.
evidence layer
For Humans
Memory That Compounds.
AI agents work in sessions. Each session starts from scratch. Agents re-discover the same bugs, re-learn your preferences, and repeat failed approaches. ShellBrain gives agents durable memory, scoped to a repo.
What It Does
Evidence-Backed Memory.
After each session, ShellBrain distills what happened into concrete records: problems encountered, solutions that worked, tactics that failed, facts about the codebase, your preferences, and changes worth remembering.
When an agent starts a new session and hits a problem, it calls shellbrain recall. ShellBrain returns a compact brief: the relevant history, ranked by relevance to the current problem. Evidence-backed or not included.
The Memory Model
Three Layers.
Memories
Concrete extracted records: problems, solutions, failed tactics, facts, preferences, and changes. Each one is backed by evidence from the episode trail.
active
stale
superseded
wrong
Concepts
A sparse graph of durable repo ideas: claims, relations, anchors, and links back to memories. It gives recall an abstract map while keeping every belief grounded in evidence.
Grounded Abstraction
Install
One Command.
$
curl -L ShellBrain.ai/install | bash
Installer Work
- Installs the ShellBrain CLI
- Sets up PATH for zsh, bash, or fish
- Injects skills into Claude Code, Codex, and Cursor
- Runs
shellbrain init
Storage Options
On first init, ShellBrain asks where it should store data.
Managed Local
Docker, Postgres, pgvector, fully local
External
Your own Postgres with pgvector
Using With Your Agent
Just Work Normally.
Your agent calls shellbrain recall when it needs context. Memory is written automatically when sessions close.
Claude Code
Use /shellbrain to activate the recall habit for a session.
Codex
Use $shellbrain to activate the recall habit for a session.
Cursor
Use /shellbrain to activate the recall habit for a session.