Context Layer · One memory for every AI agent you use

Readme

Context Layer is a personal memory that Claude, ChatGPT, Cursor, Claude Code and any other MCP-capable agent can read and write. Switch tools without re-explaining yourself: what one agent learned, the next one knows. Your context stays yours, not the vendor's.

It's open source. Run the hosted version (invite-only beta, join the waitlist) or deploy your own copy on your Cloudflare account in a few minutes. Same code, same features, either way the data is yours.

Connect an agent in one step by sharing the setup guide with it. It connects over MCP with OAuth, installs a short set of custom instructions and knows when to read and when to save. From then on every session starts with get_context and search, and ends with what's worth remembering.

Hosted or self-hosted

01 · Hosted

Use the hosted version

We run it on Cloudflare. Accounts are isolated, every write is logged, and you can export or delete everything at any time. Invite-only while the beta is small.

Join the waitlist

02 · Self-hosted

Deploy your own

Clone the repository, create the Cloudflare resources listed in the README and deploy. Your account, your Worker, your data.

git clone https://github.com/kuehlpaul/personal-context-layer.git
cd personal-context-layer && npm install
npx wrangler deploy

View on GitHub

Features

Memories

Preferences, facts, decisions and the people around you. One fact per memory, with source and date. Superseded instead of overwritten, so history stays.

Documents

Plans, specs, research and meeting notes in Markdown. Every edit is versioned with which agent wrote it, and older versions stay readable.

Lessons

When you correct an agent, the correction becomes guidance every other agent receives the moment it connects. Mistakes are made once.

Skills

Reusable, SKILL.md-style instructions. Any agent lists them, loads the one that applies to the task and follows it.

Search

Semantic search over everything you've stored, plus keyword fallback. Agents call it before working on anything personal.

Your rules

Tell it what must never be stored. Forgotten content is blocked from being re-saved. Every write by every agent is logged and editable.

Infrastructure

Protocol

Remote MCP over Streamable HTTP. Works with any MCP-capable agent, no plugin required.

Auth

OAuth 2.1 with dynamic client registration. Clients without OAuth use a personal token bound to your account.

Runtime

One Cloudflare Worker serves the MCP endpoint, the web app and this page. No servers to keep alive.

Storage

Cloudflare D1 for memories, documents, skills and the activity log. Every row is scoped to your account.

Cloudflare Vectorize with one namespace per user, embeddings from Workers AI (bge-m3).

Isolation

Per-user data isolation, independently audited. Disabling an account revokes every session, token and grant at once.

Source

Open source on GitHub. Self-host it on your own Cloudflare account, read every line that touches your data, or send a pull request.

Exit

Export everything as JSON or delete the account and all its data in one click. Nothing is kept.