Why memory
Compaction drops the exact line.
A long Grok Bot session compacts so the bot can keep working. Compaction is a summary. Summaries drop fingerprints, dates, and the exact wording of a decision.
What compaction keeps
Enough to continue: the topic, the open task, a paraphrase. That is the right trade for a context window. It is the wrong place to keep “SHA256:n4k7-Qe2m, before 14 Nov, prod key untouched.”
Retayn is the other place. The bot writes the line before the window is crushed. Later, including after compaction and in another session, it can ask for that line and quote it.
One store, more than one bot
A tag on each memory says which bot wrote it. Recall scope says who may read it: that bot, a list you choose, or every bot on the account. Claude Code, Cursor, Codex, and OpenClaw can use the same account, each with its own tag. They do not inherit a scope they were not given.
Why the record is encrypted
A store of decisions, keys, and project history is sensitive because it is useful. Retayn encrypts on the bot’s computer before upload. We cannot read the plaintext. On Grok Bot the passphrase lives on that computer: private from us, not from the host.
Research on agent memory treats an unencrypted, writable store as something an attacker can poison so later answers follow the poisoned text. Two papers we link from the homepage:
- Memory injection attacks on LLM agents (arXiv:2503.03704, March 2025).
- Memory poisoning attack and defense (arXiv:2601.05504, January 2026).
Encryption does not by itself stop a bot you trust from writing a bad memory. It does stop Retayn, and a copy of the database, from reading what was written.