Turns, memories, evidence, and the four states a memory can be in.
Memory you can question
PersonaKit treats memory as a short claim plus the words that support it, not as an invisible profile.
Turn
One message you wrote in a session. Assistant replies, tool output, pasted blocks, injected context, and PersonaKit's own extraction sessions are dropped before anything else happens.
Memory
A short statement about you (one sentence, at most 300 characters) in one of seven categories: communication, work-style, tools, preferences, background, constraints, other. For example "prefers direct, practical feedback" (fictional). A memory without evidence cannot be stored. Its id is derived from its normalised text, so the same claim found twice is one memory with more evidence.
Evidence
The quote from your own turn that supports the memory, with the session it came from.
States
| State | Meaning | Served to assistants |
|---|---|---|
unconfirmed | Proposed, not yet reviewed | Yes, labelled |
approved | You confirmed it | Yes |
restricted | You limited it | No (counted as withheld) |
archived | You hid it | No |
Only you move a memory between states. Assistants can read and propose; a proposal always starts as unconfirmed. The system never deletes a memory on its own.
Retrieval
Reads use word overlap with the question. When fewer than three memories match, approved ones fill the remaining places as standing context. This is a deliberately simple baseline, not a finished retriever: matching is on whole words in Latin letters and digits, so questions and memories in other scripts are not matched in 0.1.