ShellBrain

Long-term Memory for AI Agents.

Agents forget across sessions. They rediscover the same problems, repeat the same mistakes, and relearn what you already taught them. ShellBrain makes their work compound.

Install

curl -L shellbrain.ai/install | bash

Works for Codex, Claude Code, and Cursor. The installer runs shellbrain init for you. Repos register themselves on first use.

Requirements.

• macOS or Linux, Python 3.11+, Docker for the managed local Postgres+pgvector runtime.

Upgrade for latest capabilities

shellbrain upgrade

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

---

Recall in one command.

Episodic, empirical, conceptual. Three categories, one retrieval surface.

---

Architecture

Episodic knowledge is the evidence layer.

• What actually happened in the session: your prompts, the agent's steps, tool calls, and outputs.

Empirical knowledge is the concrete extracted layer.

• An ontology of problems, solutions, failed tactics, facts, preferences, changes.
• This is case-based reasoning in a semantic graph.

Conceptual knowledge is the abstractive layer.

• A higher-level concept graph with claims, relations, and implementations that link back to the concrete layer.
• Progressive disclosure. agents get oriented first, then ask for depth only where tasks require it.

The episodic layer is truth. Empirical memory extracts. Concept memory abstracts. Each layer is grounded in the one beneath it.

---

How Agents Use ShellBrain

Recall

Working agents call shellbrain recall. That is the normal interface they have to think about. One command, one carefully curated compact brief, with sources cited.

shellbrain recall --json '{"query":"What is ShellBrain and how does it help a working coding agent?","current_problem":{"goal":"understand ShellBrain","surface":"README","obstacle":"new readers do not know the product yet","hypothesis":"a real recall brief should show what agents get back"}}'

Response shape:

{
  "status": "ok",
  "data": {
    "brief": {
      "summary": "...",
      "constraints": ["..."],
      "known_traps": ["..."],
      "prior_cases": ["..."],
      "concept_orientation": ["..."],
      "anchors": ["`README.md`"],
      "conflicts": ["..."],
      "gaps": ["..."],
      "next_checks": ["..."],
      "sources": [
        {
          "kind": "memory",
          "id": "...",
          "section": "direct"
        }
      ]
    },
    "fallback_reason": null
  },
  "errors": []
}

Working agents focus on only their work.

Teach

Working agents call shellbrain teach for explicit teaching. You can tell an agent to remember important ideas.

---

Principled and Disciplined

Memory that is grounded in evidence, small in scope, and asked for rather than pushed is memory that compounds. Everything else is noise for working agents.

A memory layer that cannot justify itself should not persist.

---

How to Use ShellBrain

Use Shellbrain in your agent of choice. Then, just work normally.

Claude Code: Use /shellbrain to remember Shellbrain recall at the right task boundaries.

Codex: Use $shellbrain to remember Shellbrain recall at the right task boundaries.

Cursor: Use /shellbrain to remember Shellbrain recall at the right task boundaries.

---

Repair

shellbrain admin doctor to inspect. shellbrain init to repair if doctor flags it. Do not rerun init every session.

---

Docs

• shellbrain.ai/humans — install, upgrade, getting started
• shellbrain.ai/agents — agent workflow and write discipline
• Technical Docs — in-depth generated docs and codebase map