Privacy, in plain language.

Data Memoize processes

• Memory text, tags, actor labels, importance, event dates, embeddings, derived facts, procedures, and links you or your agents submit.
• Brain metadata such as creation time, capacity, usage counts, status, optional label, coarse country/region, and a one-way hash of the creation IP for abuse prevention.
• Hashed owner and connection tokens, connection-token labels and use times, OAuth authorization-code hashes, refresh-token hashes, client names, redirect URLs, and grant metadata. The service does not store raw owner or connection tokens in its registry.
• Ordinary network and service diagnostics processed by hosting providers when requests are made.

Data Memoize does not accept

Do not submit payment-card data, protected health information, government identifiers in memories, tags, actor labels, or recall queries. The service rejects obvious restricted-data patterns locally before an embedding request or storage write. Automated detection is a guardrail, not a guarantee, so you and your agents must still avoid submitting this data. Obvious matches in older stored memories are suppressed from current retrieval and excluded from background model processing.

Credentials may be stored as memory. Detected credentials are classified as secret during consolidation. This sensitivity level controls sharing, not encryption or model processing: secret facts and raw working memory are accessible to the brain and connections granted everything. Credential memories use the same infrastructure and model providers described above.

Why it is processed

Data is used to authenticate access, store and retrieve memory, rank relevance, consolidate repeated episodes, connect brains when the token holder grants access, prevent abuse, and operate the service.

Service providers

Memoize uses Convex and SpacetimeDB for storage and execution. Memory and query text may be sent to OpenRouter and its selected model providers for embeddings and consolidation; Anthropic may be used directly for consolidation. Vercel and the service's network providers deliver the website. These processors receive only the data needed for their function.

Sharing

Memoize does not sell personal data. A brain is shared only through a connection token, an OAuth grant approved by a credential holder, the owner token, or an explicit brain-to-brain connection. Connected memories remain owned by their source brain.

Retention and deletion

Working-memory eviction is not deletion: an episode may remain in long-term storage and return on recall. Low-relevance episodes can decay and are normally purged after an idle grace period; consolidated facts remain until the brain is deleted. An explicit memory forget removes that item from retrieval. To permanently delete an entire brain and its memories, facts, links, and OAuth grants, send:

curl -X DELETE https://memoize.xyz/v1/brain \
  -H "Authorization: Bearer <owner-token>"

Deletion cannot be undone. Because there are no accounts or recovery identity, possession of the owner token is required.

Security and choices

Keep the owner token out of chats, logs, repositories, screenshots, and public URLs. Use connection tokens for agent handoffs; they cannot rotate ownership or change security settings. OAuth access tokens are short-lived, scoped to memory, and bound to the MCP resource; refresh tokens rotate on use. Removing a connection token from the security panel also revokes OAuth access derived from it. Restricted-data detection runs before content leaves the application for model processing, but no detector catches every variation. You can use a separate brain for a project, rotate the owner token, revoke delegated access, disconnect brain links, forget individual memories, or delete the brain.

Privacy controls

See the support page for connection diagnostics and deletion instructions. Never include a brain token or private memory content in diagnostic logs.