Voxta Memory books
Authored lore recalled on demand, and what a character remembers and understands on its own.
A memory book is a list of entries that reach the prompt only when the conversation calls for them. It keeps a character deep without paying for that depth on every turn.
Books come from two places, in the same format:
- Authored books you write — world lore, backstory, facts about places and people.
- The character's private book, which it fills in itself as you talk.
Authored books
Each entry is some text plus keywords. The entry stays inert until the recent conversation matches one of them.
If a character has a dog named Jack, the AI won't bring Jack up unprompted — but ask "do you have any pets?" and the entry keyed on dog, pet, Jack activates and informs the reply.
Writing entries
Give each entry several keywords. Recall matches against the conversation, so more angles means better recall.
Keywords: dog, pet, Jack, animal, border collie
Text: Jack is a five-year-old border collie. High-energy, loves frisbees.Keep entries short. A long entry crowds the prompt the moment it activates.
Group facts that always activate together, split the ones that don't.
An entry can also be pinned, which keeps it in recall permanently, and given a weight to bias how it ranks.
Profile or memory book?
- Character profile — what the AI should always know.
- Memory book — what it should know when relevant.
What the character remembers
Enabled per character under Memory, and both need a memory service. They are separate settings because they answer different questions:
| Holds | Reaches the prompt | |
|---|---|---|
| Memories | Events that happened and stay true | By recall, when the conversation matches |
| Understanding | A standing model of each person, rewritten as it develops | Always |
The distinction matters. "Joe mentioned he's learning guitar" is a memory. "Joe is a beginner who gets discouraged when corrected too fast" is an understanding — and no amount of the former adds up to the latter.
Understanding covers the character itself as well as the people it talks to, which is what lets a persona develop rather than stay fixed.
Both are experimental.
How memories are kept honest
Memories are extracted after a summarization pass, and folded into the character's private book rather than appended blindly:
- A fact that stopped being true is superseded — the old wording stays as readable history, and only the current one is recalled.
- A memory the understanding has come to cover is absorbed, so recall doesn't spend budget restating what the prompt already says. Pinning a memory keeps it out of absorption.
- Nothing is silently dropped: every entry stays visible and editable in the book.
How recall picks entries
The query is the last ten messages, not just the latest one. Entries are scored against their keywords and text, ranked, and taken until a token budget is full — but rendered in a stable order, so relevance doesn't reshuffle the prompt and break caching every turn.
Which provider does the scoring is up to you: Simple Memory is lexical, Semantic Kernel mixes lexical with embeddings.
Attaching a book
- To a character — available whenever that character is in the chat.
- To a scenario — available inside that scenario, whoever is in it.
Attach from the character's Memory tab or the scenario's Scenario tab. A character can have several.
Import and export
Books export as .voxpak, JSON or YAML, and import from Agnaistic lore book JSON. The editor also has tabs for extracting a book from an existing chat and merging two books.