Your local-first memory layer for AI chats, coding agents, and project context.
Landing Page · npm · Install · Technical FAQ · Download Stats · Roadmap
No backend. No accounts. No tracking.
AI work is powerful, but context is fragmented. Switch models, hit token limits, change accounts, switch platforms, switch coding agents — and the decisions, attempts, and threads evaporate. Your code lands in Git. Your context does not.
ContextVault fixes this by making browser conversations and terminal work sessions portable, local, and reusable. Git tracks code. ContextVault tracks context.
Download the latest installer for your platform:
| Platform | Download |
|---|---|
| Windows | Download (.exe) |
| Linux | Download (.AppImage) |
| macOS | Build from source (npm run package:mac) |
Windows and Linux downloads resolve to the latest published GitHub release assets. The app works offline. macOS currently requires a local source build (no Apple Developer code-signing setup yet).
Run immediately without installing:
npx @aliabdm/contextvault initOr install globally:
npm install -g @aliabdm/contextvault
contextvault initThe npm package provides Vault Terminal and the Unified Context Engine. It does not replace the browser extension.
npm install
npm run buildThen load dist/ into Chrome via chrome://extensions → Developer Mode → Load unpacked.
- Add a project — choose a local folder. Add more later and switch from the sidebar.
- Start recording — name the session and choose Codex, Claude Code, Cursor, Terminal, or another source.
- Work visually — use filters and result screens; sessions written by CLI or integrated agents appear live.
The Record button launches the bundled contextvault record process. Desktop never watches your screen or terminal in the background. Automatic classification runs on-device — no model, no API call, no telemetry.
 |
 |
npx @aliabdm/contextvault init
contextvault recordInside the recorder:
/source codex
/title Fix auth middleware
/user The login redirect is broken.
/agent I found the issue in middleware order.
/decision Keep auth checks in middleware and policy checks in controllers.
/task Add regression test for redirect loop.
/problem Session cookie is missing on callback.
/end
Other commands:
contextvault list # list saved sessions
contextvault search "auth" # search across sessions
contextvault history --since 2w # project history, last 2 weeks
contextvault decisions auth --source codex # Codex decisions about auth
contextvault problems redis --since 30d # Redis problems, last 30 days
contextvault retrieve "auth middleware" # ranked evidence retrieval
contextvault prepare "auth middleware" # agent-ready context packageSupported filters: --type, --source, --since (24h, 14d, 2w, or ISO date), --limit.
|
|
|
|
|
|
|
| Feature | Description |
|---|---|
| Desktop App | Visual context manager for Windows and Linux; macOS source build. Multi-project support, live watched-file sync, dedicated GUI screens for every package command. |
| Browser Capture | Records conversations from ChatGPT, Claude, Gemini, Perplexity, Poe, DeepSeek, Copilot. Hybrid DOM + network engine. Exports as Markdown or ZIP. |
| Terminal Capture | Records human, AI, and coding-agent sessions from the terminal. Raw text preserved exactly — no summarization, no context loss. |
| Unified Context Engine | Imports browser exports, normalizes both sources into a shared model, indexes locally, and retrieves cross-surface evidence. |
| Local-first storage | Browser chats in IndexedDB; terminal sessions in .contextvault/. No backend, no cloud sync. |
| Deterministic retrieval | Phrase + token + type + recency ranking. No embeddings, no external API. Same query, same vault, same result every time. |
| Focused views | Dedicated History, Decisions, Problems, Tasks, Retrieve, Search, and Prepare screens — in Desktop and CLI. |
| Prepared context | Generates a focused Markdown package ready for the next Codex, Claude Code, Cursor, or other agent. |
| Project memory | Auto-generated decisions/tasks/problems block in memory.md. Manual content preserved. |
| Session links | Explicit relationships between sessions with a reason label. |
| Timeline export | Chronological project-context timeline as Markdown. |
| Private by default | No backend, no accounts, no telemetry, no external AI API calls. |
Browser Capture Terminal Capture Desktop App
DOM Observer + Network Monitor Vault Terminal CLI Electron UI
| | |
Stream Assembler Raw Context Events Preload API
| | |
Capture Engine Markdown Sessions Main Process
| | |
Background Service Worker .contextvault/ Context Engine
| | |
IndexedDB Storage Local Filesystem Storage Local Filesystem
| | |
Markdown / ZIP Export Sessions Visual Output
\ / /
\ / /
--------------- Shared Context Layer ---------------------------
|
Context Engine
Normalize / Index / Retrieve
Prepare / Memory / Link
|
Markdown / ZIP / Timeline
Key ideas:
- Browser: DOM is the source of truth; network is an enhancement layer; the Stream Assembler builds final messages.
- Terminal: Raw input is the source of truth; Markdown files are durable memory; no summarization means no context loss.
- Desktop: A GUI over the same
.contextvaultfiles — not a second recorder, not a parallel database. Sessions written from the terminal appear in Desktop automatically, and vice versa.
.contextvault/
config.json
memory.md
links.json
imports/
browser/
index/
context-index.json
sessions/
exports/
prepared-context.md
context-timeline.md
contextvault-terminal-export.md
| No | Yes |
|---|---|
| No backend | Everything stays local |
| No accounts | You control your own context |
| No telemetry | Export when you choose |
| No external AI APIs | Terminal capture works without calling AI services |
| No automatic cloud sync | Local files and local browser storage by default |
The .contextvault/ folder is git-ignored by default because terminal sessions may contain raw private context, paths, prompts, logs, secrets, or customer information.
The public download stats page reads GitHub release asset counts directly from the browser — no device ID, no fingerprint, no background ping from the Desktop app. It counts download events, not unique people.
npm install
npm run build # extension build
npm test # run test suiteLoad dist/ into Chrome as an unpacked extension.
cd desktop
npm install
npm run dev # electron-vite dev
npm run package:win # build Windows installer
npm run package:mac # build macOS app
npm run package:linux # build Linux AppImagedocker compose up dev # extension dev server on :5173
docker compose --profile testing up dev test-server # + ChatGPT mock on :8080- Published
@aliabdm/contextvaultnpm CLI - Desktop App (Windows/Linux installers; macOS source build)
- Package-powered Desktop GUI with every
contextvaultcommand and live external-session refresh - Browser Capture (ChatGPT, Claude, Gemini, Perplexity, Poe, DeepSeek, Copilot)
- Terminal Capture (Codex, Claude Code, Cursor, human notes)
- Unified Context Engine: normalize, index, retrieve, prepare
- Markdown, ZIP, and directory browser import with duplicate prevention
- Focused task, decision, and problem views
- Filtered history and evidence queries by source, type, and time
- Prepared agent context packages
- Generated project memory block (manual content preserved)
- Explicit session links
- Context timeline export
- Multi-project switching with live watched-file sync
- Auto-tagging system
- Retrieval ranking improvements and optional local semantic indexing
- Automatic context linking suggestions
- Memory conflict and stale-task detection
- Optional local-model or user-configured provider adapter for grounded natural-language answers
- MCP server
- VS Code extension
- Claude Code, Codex, and Cursor direct integrations
- Encrypted backups
- Optional self-hosted sync
- Context Repositories (next phase): push a vault to a remote repository, invite collaborators, pull shared context, review diffs, edit locally, and push updates with Git-like history and conflict handling
- Provider UI changes may require selector updates.
- Terminal capture records explicitly entered context; it does not automatically intercept shell processes.
- Context Engine retrieval is lexical and deterministic; it does not use embeddings or semantic model calls.
- Browser conversations enter the engine through explicit Markdown/ZIP import; automatic sync from IndexedDB is not implemented.
- Natural-language answer generation is not built in; current commands retrieve grounded evidence and prepare it for an agent.
- Browser storage is local to the browser profile; terminal storage is local to the project where
.contextvault/is initialized.
Mohammad Ali Abdul Wahed
| GitHub | aliabdm |
| Mohammad Ali Abdul Wahed | |
| X (Twitter) | @Maliano63717738 |
| Portfolio | senior-mohammad-ali.vercel.app |
| Medium | @aliabdm |
| Dev.to | @maliano63717738 |
Dev.to:
- The Code Was in Git. The AI Conversations to Implement It Were Gone
- ContextVault: Own Your AI Context Across Models, Agents, and Time
- Your AI Chats Are Locked in Corporate Silos — I Built an Open-Source Vault to Rescue Them
Medium:
- The Code Was in Git. The AI Conversations to Implement It Were Gone
- ContextVault: Own Your AI Context Across Models, Agents, and Time
- Your AI Chats Are Locked in Corporate Silos. I Built an Open-Source Vault to Rescue Them.
- GitHub Discussions — ask questions, share workflows
- GitHub Issues — bug reports and feature requests