cli-memory-bridge

cli-memory-bridge is a local-first memory bridge for terminal coding agents and MCP-enabled CLIs. It imports conversation history from supported providers on your machine, stores canonical transcripts and embeddings locally, and serves that memory back through cmb commands and an MCP-compatible stdio server.

The current Rust build supports transcript import and retrieval for:

• Codex
• Claude
• Gemini
• Copilot
• Zed
• OpenCode
• Hermes
• Antigravity CLI

It also ships an npm package surface so MCP configs can launch it through:

npx -y @aminovsky/cli-memory serve

Features

• cmb init to detect providers and import local conversations
• cmb refresh to incrementally import only new or changed sources
• cmb resume <hash-id> to restore a conversation transcript by stable hash
• cmb conv-search <query> to search imported conversation content
• cmb forget <provider> <hash-id> to soft-ban a conversation from future retrieval
• cmb serve to expose the stdio MCP tool surface
• cmb doctor and cmb stats for diagnostics
• cmb install <provider> to render provider install/config bundles
• cmb unlink <provider> and cmb unlink --all to render provider cleanup bundles
• cmb uninstall to render full package removal guidance

Install

npm / npx

For MCP usage and command execution through npm:

npx -y @aminovsky/cli-memory --help

On first install, the npm package downloads a platform-specific cmb binary from the matching GitHub release for this package version and stores it under npm/vendor/<platform>/.

If you install the package globally:

npm install -g @aminovsky/cli-memory
cmb --help

Local Rust build

cargo build --release
./target/release/cmb --help

Quick start

Import local transcripts and start using retrieval:

cmb init
cmb doctor
cmb stats
cmb conv-search "run the app"

Run the MCP server through npx:

npx -y @aminovsky/cli-memory serve

MCP launch

The npm package is designed for MCP configs that accept command plus args. The preferred launcher shape is:

{
  "command": "npx",
  "args": ["-y", "@aminovsky/cli-memory", "serve"]
}

Provider install

Render a provider-specific install bundle:

cmb install gemini
cmb install zed
cmb install codex

Current install behavior is render-only. It prints the config/assets you should apply for that provider.

Provider unlink

Render cleanup instructions for one provider:

cmb unlink gemini
cmb unlink zed
cmb unlink codex

Render cleanup bundles for every supported provider:

cmb unlink --all

unlink removes provider wiring only. It does not remove the cli-memory package or the local memory database.

Full uninstall

Render full uninstall guidance:

cmb uninstall

The intended order is:

cmb unlink --all
npm uninstall -g @aminovsky/cli-memory

If you only use npx -y @aminovsky/cli-memory ..., there may be no global npm package to remove.

Supported local transcript sources

• Codex: ~/.codex/sessions, ~/.codex/session_index.jsonl
• Claude: ~/.claude/projects
• Gemini: ~/.gemini/tmp/*/chats/session-*
• Copilot: ~/.copilot/session-state/*/events.jsonl
• Zed: ~/.config/zed/conversations/*.zed.json
• OpenCode: ~/.local/share/opencode/storage/session_diff/*.json with opencode.db fallback
• Hermes: ~/.hermes/sessions / ~/.hermes/history
• Antigravity CLI: ~/.gemini/antigravity/brain/*

Release notes

For npm release prep, this repo now includes:

• package metadata in package.json
• Apache 2.0 license in LICENSE
• npm bin entrypoint in npm/bin/cli-memory.js
• npm postinstall downloader in npm/install.js
• shared platform mapping in npm/platform.js

Before publishing to npm:

• build release binaries for each supported platform
• upload them to the GitHub release tagged v<package.json version>
• use asset names in the form cmb-<version>-<platform-arch> and cmb.exe-<version>-<platform-arch> for Windows
• verify with npm pack --dry-run and a clean npx -y @aminovsky/cli-memory --help install test

Status

This build is production-oriented but still evolving:

• transcript import and retrieval are implemented
• provider lifecycle commands are implemented
• some providers are reconstructed from artifact trails rather than full native session models
• graph/knowledge-linking is intentionally deferred

Special Thanks

This project takes reference and inspiration from:

• rohitg00/agentmemory
• aiming-lab/SimpleMem
• RyanCodrai/turbovec