Beside, in one page

Beside is a local-first AI memory layer for your computer. It silently captures what you do across apps, indexes that activity into a self-organising Markdown wiki, and exposes the result to every AI agent you already use through the Model Context Protocol (MCP).

The product runs entirely on your machine. Every line of code is open source under MIT, the model can be local (Ollama) or hosted (OpenAI / OpenAI-compatible), and your raw screenshots, audio, and text never leave your disk unless you explicitly export them.

What you get out of the box

• A timeline of your work, built from screenshots, focused windows, URLs, idle/active state, and (optionally) audio with on-device transcription.
• A self-organising wiki at ~/.beside/index/ and a stable mirror at ~/.beside/export/markdown/ — readable, greppable, and diff-friendly.
• Sessions, meetings, and day events, automatically derived from raw activity so the agent gets durable nouns, not just frames.
• Semantic search over both raw frames and the higher-level memory chunks, via local embeddings.
• An MCP server on 127.0.0.1:3456 that any compatible agent — Claude Desktop, Cursor, Windsurf, ChatGPT desktop with MCP — can call into.
• A CLI (beside) for status, doctor, init, capture, indexing, MCP, and resets.
• A native macOS desktop app with packaging, auto-update, permissions, and a renderer that can mount custom hook widgets.
• A plugin system that lets you replace any capture, storage, model, index, hook, or export behaviour without forking the product.

How it’s positioned

Most AI products try to convince you to move your work into them. Beside does the opposite: it sits beside whatever you already use and turns it into context your AI can recall. That makes it useful on day one, and durable as your toolchain changes — the memory is yours, in plain Markdown plus SQLite, on your own disk.

The unique selling points worth keeping in mind as you read these docs:

1. Local-first by construction. Capture, embeddings, indexing, and the MCP server all run on-device. Hosted models are opt-in.
2. Plain Markdown wiki. No proprietary format. You can read it in any editor, version it in git, or wipe it and rebuild it from raw events.
3. Replaceable layers. Capture, storage, model, index, hooks, export — each one is a typed interface (ICapture, IStorage, IModelAdapter, IIndexStrategy, IHookPlugin, IExport) backed by drop-in plugins.
4. MCP-native. Beside is not a chat UI. The product surface is your existing agent, augmented with persistent memory.

Where to go from here