Measured performance
Published at ACL 2026, Hindsight — the memory engine Savantis builds on — reached 83.6% accuracy on LongMemEval and 83.2% on LoCoMo with an open 20B model.
Read the ACL publicationSavantis Memory
One memory, shared by every assistant you use.
You tell it once. Every assistant you connect can read it, correct it, and build on it. Nothing to run, nothing to host — it plugs into your assistants through MCP.
$9.99 per month, taxes extra · 5-day trial · cancel any time
Three things, and that is the whole product
Not a database you administer. Three ideas you already think in.
- A memory
- Something worth keeping from a conversation. Your assistant writes it as you talk, or you write it yourself.
- A space
- A grouping the way you actually live: Personal, Work, House, Music. Every account has one Personal space that cannot be archived, so nothing is ever homeless.
- A confirmed reference
- A fact you approved yourself. These come back before anything a model inferred, they keep their source and the date you confirmed them, and every version stays readable.
What the research says
Memory improves continuity. Control keeps people in charge.
Memory does not make an AI safe by itself. It adds capability and context; safety depends on how memory and agency are governed.
Safety by design
Yoshua Bengio and his coauthors separate intelligence — learning, memory, and reasoning — from the ability to act and pursue goals. Their direction is to preserve the first while constraining the other two.
Read the Bengio and LawZero publicationSavantis turns that distinction into controls you can use: memories stay readable, correctable, exportable, and deletable; confirmed references come first; each connected client can be revoked on its own.
What you can do with it
- Add, recall, correct, delete
- From any connected client, in the conversation itself.
- Connect any MCP client
- One public endpoint at context.savantisai.com/mcp. Claude, ChatGPT, command-line agents, or a Hermes running on your own machine. Each client asks for your approval once and can be revoked on its own.
- See contradictions instead of losing them
- When a new fact disagrees with an older one, both stay visible with their dates, and you pick which one holds.
- Import and export
- Bring in what you already wrote, and take everything out whenever you want, in a plain portable file.
- Reading costs nothing extra
- Recall, search, organizing, confirming and importing use storage and local retrieval only. They invoke no inference and debit no tokens.
Your memory stays yours
What you get, in plain terms.
- One account, one memory
- Your memory answers to your account, and the account comes from how you signed in.
- You approve every client
- Each assistant you connect asks for your approval once, and you can cut it off on its own without touching the others.
- You sign in the way you already do
- Google, Microsoft, or a six-digit code sent to your address. Nothing new to remember.
- You decide what it keeps
- It remembers what is written to it, and you can read every entry with its date and where it came from.
- You can take it back
- Export everything whenever you want, and delete what you want deleted.
Good to know
Two things, written in the terms as well.
- Savantis Memory is in beta, and built for personal use.
- Export is always available, so what you put in stays yours to take.
Say it once.
5 days to find out whether it changes anything. A payment method is collected at purchase and nothing is charged during the trial.
Start the 5-day trialBeta · personal use