Skip to main content
Everything in one Memory Crystal account belongs to that account. Agent, channel, session, and project settings change what a request returns. They do not create a separate account, and they are not a security boundary between tenants. If two parties must not be able to read each other’s data, give them separate Memory Crystal accounts and API keys.

​
Account

An API key authenticates as one user. Reads and writes use that user’s rows. A memory id from another account is not found. Hiding a row inside the account (a knowledge base the agent is not on, a channel the request did not name) also returns not found, so a caller cannot tell “missing” from “not visible.”

​
Agent identity

agentId is a string you declare, usually the platform slug the installer writes (claude-code, codex, cursor, and the others). Hooks resolve it in this order: hook input, MEMORY_CRYSTAL_AGENT_ID, the known host identity, then ~/.memory-crystal/config.json. Knowledge-base row access uses that id: An inactive base is hidden from everyone. Listing bases with no agentId and no channel is the management list: the filter is off, so you can see bases this agent cannot query. Pass agentId or channel and the filter turns on. Add a slug with crystal_set_knowledge_base_access (action=add) or in the dashboard under Knowledge → Agents.

​
Chunks are stricter than the row

Seeing the knowledge base does not mean seeing every chunk.
  • A chunk with no channel is shared with everyone who can see the base.
  • A chunk with a channel uses the channel rules below. A request with no channel does not receive it; a scoped request may also see chunks in its bare prefix or suffix scope, except for group and numeric-peer channels.

​
Channel

A channel is a scope string on a memory or a request, often platform:workspace from the hook. It partitions non-KB memories inside the account. If your memories seem to have disappeared, check the request channel: a scoped channel hides global (channel-less) memories, while an ordinary bare channel includes them. Keep the same intended scope when saving and reading. Knowledge-base row checks ignore the channel. Chunk checks do not.

​
Session and project

These are filters on one account’s data. They do not isolate that data from another key on the same account.

​
What partitioning does not do

  • It does not stop a person who holds the account’s API key from asking without a channel or with a management list.
  • It does not stop a second agent you install on the same account from seeing channel-less memories and * knowledge bases.
  • It does not apply across accounts. Another account’s key sees nothing here.
Separate accounts are the privacy boundary.