A portable memory layer for LLM conversations.

Every time you start a new LLM chat, all context is lost. llm-brain fixes that. It gives your LLM a persistent, structured state file that accumulates decisions, progress, goals, and facts across sessions.

At the end of a conversation, you ask your LLM to generate a brain state snapshot. llm-brain save merges it into a growing state file. At the start of any new conversation, llm-brain load copies the full context to your clipboard. Paste it in and the LLM instantly has complete continuity.

End of chat  ->  Ask LLM to generate brain state  ->  llm-brain save
New chat     ->  llm-brain load  ->  Paste  ->  LLM remembers everything

• Node.js >= 18.0.0

Install globally to make the llm-brain command available system-wide:

npm install -g llm-brain

Note: Global installation (-g) is required for the CLI. A local install (npm install llm-brain) will not register the command on your PATH. If you prefer not to install globally, you can use npx llm-brain instead.

cd my-project
llm-brain init

This creates a .llm-brain/ directory with the following structure:

.llm-brain/
  brain.json         -- The persistent state file
  config.json        -- Configuration preferences
  history/           -- Snapshot archive of each session
  templates/         -- Prompt templates

llm-brain template show extract

Copy this prompt. At the end of any LLM conversation, paste it in and ask the LLM to generate your brain state as JSON.

After your LLM generates the JSON, copy it to your clipboard, then run:

llm-brain save

Or pipe it directly:

llm-brain save --from stdin < state.json

llm-brain load

Your full context is now on your clipboard. Paste it at the start of your new chat and the LLM will have complete context to continue where you left off.

Initialize .llm-brain/ in the current project directory.

llm-brain init

Save and merge a new brain state.

llm-brain save                                    # Read from clipboard (default)
llm-brain save --from stdin                       # Read from piped stdin
llm-brain save --from file --file ./state.json    # Read from a file
llm-brain save --replace                          # Replace instead of merging

Load the brain state for a new conversation.

llm-brain load                      # Copy to clipboard (default)
llm-brain load --to stdout          # Print to terminal
llm-brain load --session 3          # Load from session #3

Display the current brain state in a readable format.

llm-brain show                      # Pretty-printed view
llm-brain show --json               # Raw JSON
llm-brain show --compact            # One-line summary
llm-brain show --section decisions  # Show a specific section

Available sections: facts, decisions, activeState, keyFiles, knownIssues, preferences, nextSteps

List all saved sessions.

llm-brain history                   # Show last 10 sessions
llm-brain history --limit 5         # Show last 5 sessions

Reset the brain state.

llm-brain reset                     # Full reset (prompts for confirmation)
llm-brain reset --keep-config       # Reset state only, keep config
llm-brain reset --yes               # Skip confirmation prompt

Manage prompt templates.

llm-brain template show extract     # Show the extraction prompt
llm-brain template show inject      # Show the injection prompt
llm-brain template reset            # Reset templates to defaults

The brain.json file is a structured JSON document with seven layers:

{
  "version": "1.0.1",
  "project": "my-project",
  "sessionCount": 7,

  // FACTS -- project truths that accumulate over time
  "facts": {
    "stack": ["TypeScript", "Express", "PostgreSQL"],
    "architecture": "Monolith",
    "environment": { "os": "macOS", "node": "20.x" },
    "custom": ["API on port 3000", "ESM modules"]
  },

  // DECISIONS -- what was decided and why
  "decisions": [
    { "what": "Use PostgreSQL", "why": "Relational data fits", "when": "2026-02-15", "reversible": true }
  ],

  // ACTIVE STATE -- what is happening right now
  "activeState": {
    "currentGoal": "Add WebSocket support",
    "inProgress": ["WebSocket handler"],
    "blockers": ["WS drops after 30s idle"],
    "recentlyCompleted": ["Error handling", "Zod validation"]
  },

  // KEY FILES -- important file map
  "keyFiles": [
    { "path": "src/index.ts", "role": "Main server entry" }
  ],

  // KNOWN ISSUES -- bugs and workarounds
  "knownIssues": [
    { "issue": "pg ESM import fails", "workaround": "use createRequire", "severity": "low" }
  ],

  // PREFERENCES -- how the user prefers to work
  "preferences": {
    "style": "Learning by doing",
    "constraints": ["$0 cost only"],
    "communication": "Explain reasoning, not just code"
  },

  // NEXT STEPS -- ordered priority queue
  "nextSteps": ["Fix WS timeout", "Add rate limiting"]
}

llm-brain does not blindly overwrite. It intelligently merges each new session with the existing state:

Edit .llm-brain/config.json to customize behavior:

{
  "saveSource": "clipboard",       // "clipboard" | "stdin" | "file"
  "loadTarget": "clipboard",       // "clipboard" | "stdout"
  "maxHistory": 50,                // max session snapshots to keep
  "gitIgnore": false,              // false = commit brain state (recommended)
  "autoDetectProject": true        // auto-detect name from package.json or git
}

• Commit your .llm-brain/ directory. It contains only JSON and markdown. Version-controlling it alongside your code is recommended.
• Works with any LLM. Compatible with ChatGPT, Claude, Gemini, GitHub Copilot, Cursor, and any tool that accepts text input.
• The extraction prompt is critical. Run llm-brain template show extract and paste it at the end of every session before closing.
• Sessions are cumulative. Each save builds on the previous state. The brain becomes more complete over time.

MIT