← LLMemory

Privacy Policy

Last updated: 17 August 2026.

Summary

LLMemory is local-first. Your conversations live in a database on your own device and, if you enable sync, in storage you own and control (your Google Drive, or optionally a private GitHub repository). LLMemory's remote services do not persist normal users' conversation content. A narrow directory-review exception uses intentionally synthetic sample conversations hosted by LLMemory so OpenAI and Anthropic can test the connector without a Google account. That fixture contains no real user conversation data. We do not sell data, show ads, or use your data to train models.

This policy covers every LLMemory surface: the browser extension, the llmemory-cli command-line tool and its local MCP server, the web app at app.llmemory.xyz, and the LLMemory connector for AI assistants (ChatGPT, Claude, and other MCP clients).

Data we collect

• Conversation content: the text of AI chats you choose to capture with the browser extension (on supported AI sites, only content already visible in your own logged-in session) or import with the CLI from local tools such as Claude Code and Codex.
• Conversation metadata: source (which AI tool), titles, timestamps, and conversation identifiers, used to organize and search your vault.
• Sign-in data: when you connect Google Drive, OAuth tokens scoped to drive.file, which limits LLMemory to files it created itself. We never read your other Drive files, your email, contacts, or profile beyond what the sign-in flow requires. The optional GitHub connection is likewise limited to the vault repository.
• Preferences: settings such as theme and capture on/off, stored locally in your browser or on your machine.
• Directory-review fixture and sign-in records: OpenAI and Anthropic reviewers can use dedicated LLMemory demo accounts containing only synthetic sample conversations. LLMemory stores the synthetic fixture plus the reviewer authentication and OAuth records needed to make those test accounts work. These records are isolated from real user vaults.

We do not collect browsing history, location, financial, health, or contact data, and we run no advertising or third-party tracking.

How we use it (purposes)

• Backing up your conversations to storage you own.
• Building a search index so you can find past conversations.
• Answering your own requests, such as exports, searches, and the assistant tool calls described below.
• Providing isolated synthetic test data so connector-directory reviewers can reproduce the submitted tool tests without accessing a real user's vault.

That is the whole list. No advertising, no profiling, no model training, no sale of data.

The AI-assistant connector (tool inputs and outputs)

When you add LLMemory as a connector inside an AI assistant, the public connector surface contains nine read-only tools. Five retrieve or summarize conversation history:

• list_sources: input none; output which AI tools have conversations in your vault, with counts.
• list_recent: input an optional limit and source filter; output titles, sources, dates, and IDs of recent conversations.
• search_chats: input your search query plus optional search filters; output matching conversation titles, snippets, and IDs.
• get_conversation: input a conversation ID plus optional format; output that conversation's transcript.
• get_context_pack: input a conversation ID or source plus an optional limit; output bounded context from the requested conversation history for resuming work.

Four additional read-only helper/setup tools are also available:

• whoami: input none; output source counts and a small set of recent conversation metadata so the assistant can orient to the connected vault.
• list_skills: input none; output names and descriptions of packaged LLMemory instruction skills. This does not read conversation content.
• get_skill: input a packaged skill name; output that static LLMemory instruction text. This does not read conversation content.
• llmemory_setup: input an optional setup step; output vault-readiness status or static setup/instruction guidance.

Verified Google Drive accounts can also have four read-only, alpha-gated project-continuity tools registered. Directory-review demo accounts do not receive these tools:

• list_projects: input none; output cloud-eligible project metadata and assignment counts already stored in the LLMemory vault.
• get_project: input a project ID; output that project's metadata plus bounded summaries of assigned conversations.
• search_project_chats: input a project ID, query, and optional search filters; output matching snippets only from conversations assigned to that project.
• get_project_context_pack: input a project ID plus optional limits; output a bounded, source-linked context pack from that project's assigned conversations.

These tools run only when you (or the assistant acting on your request) invoke them, and none of them can modify or delete anything. Tool outputs become part of your conversation in the assistant you are using, where they are handled under that provider's own privacy policy (for example OpenAI's for ChatGPT, Anthropic's for Claude).

Who receives your data (recipients)

• Your own storage: Google Drive (and optionally GitHub) accounts that you connect. Google and GitHub process this data as your providers, under their terms with you.
• The AI assistant you invoke tools from: tool outputs listed above are returned into that assistant session at your request.
• Cloudflare: hosts this website and the two small services below, as our infrastructure provider. It also hosts the synthetic directory-review fixture and the authentication records used by those test accounts.

Nobody else. No data brokers, no advertisers, no analytics vendors.

What our servers store, and for how long (retention)

Normal-user conversation content is not persisted on LLMemory's remote MCP service. Sync moves that data between your device and your own storage over HTTPS. The narrow directory-review fixture described above is synthetic and is the only conversation-like content intentionally hosted by the connector service. Two small services exist to make sign-in and connector access work:

• OAuth token exchange (api.llmemory.xyz): completes the Google sign-in handshake so the client secret stays server-side. It holds short-lived authorization codes for minutes, until they are used or expire, and retains nothing else.
• Connector service (mcp.llmemory.xyz): serves the assistant tools above. For normal users, it stores the OAuth client registrations and access grants needed to keep your connector signed in, until you disconnect or revoke them; conversation content passes through transiently to answer a request and is not retained. For directory-review accounts, it also retains the synthetic fixture and the reviewer account verifier/OAuth records while those review accounts remain enabled. Reviewer credentials and grants can be rotated or revoked by LLMemory.

Everything else is retained only as long as you keep it: your vault lives in your Drive (or local database) until you delete it, and the local search index is removed when you delete the vault or uninstall.

Your controls

• Export your entire vault at any time as plain JSON and Markdown (llmemory export-all or the web app).
• Delete any conversation or the whole vault by deleting the files in your Drive or the local database. LLMemory does not keep a second hosted copy of normal-user conversation content.
• Disconnect Google Drive from the extension or CLI, or revoke LLMemory's access directly at myaccount.google.com/permissions.
• Remove the connector from your assistant's settings (for example ChatGPT's connector settings), which revokes its access grant.
• Uninstall the extension or CLI, which stops all capture and sync.

Changes to this policy

If we change this policy, we will update this page and the date at the top. Material changes (for example, any new recipient of data) will be called out in the product before they apply to you.

Contact

Questions: hello@llmemory.xyz or llmemory.xyz/contact