nua
Memory that reads the room

How do you give an AI the full picture without leaking a private chat?

AI memory was built for 1:1 chat. In a group, the best reply depends on who's there and what each of them knows. To stay safe, most setups make sure private and group chats don't mix, leaving the AI to answer without the full picture.

We've been working on this at Nua. We built a way to track who was there when something was said, so the AI can draw on that context without giving anyone away. It works, and we haven't seen anyone else do it. Now we want to shape it with teams building multiplayer AI who are wrestling with these same trade-offs.

An example with invented people

Maya asks what to cook. The AI steers clear of soy without saying why.

  • JJess
  • MMaya
  • SSam

Three friends who share one AI assistant, in a group chat and in chats of their own. Most of what they tell it is for everyone. One thing isn't.

  1. Monday
    JAIJess and the AIprivate

    JessEnded up in the ER after Maya's birthday dinner 😩 Turns out I'm allergic to soy. Please don't tell her, she'd feel awful

    AII'm so sorry, Jess. Glad you're okay. That stays between us.

    Jess is allergic to soy. She found out in the ER, after Maya's birthday dinner

    Heard by Jess
  2. Tuesday
    JMSAIGroup chatall three

    SamOfficially vegetarian as of today 🥦

    MayaProud of you!! I'll cook accordingly 😌

    AINoted! 🌱

    Sam is vegetarian

    Heard by all three
Thursday · group chat

Maya asks what to cook on Friday. The AI knows something she doesn't. If it forgets, it might suggest ramen. If it explains, Jess's secret is out.

JMSAIGroup chatall three

MayaHosting Friday! Ramen night or tacos? 🍜🌮

AI · with NuaTacos!1 Set it up as a build-your-own bar so everyone makes theirs the way they like, with black beans for Sam2. Want a shopping list?

  1. 1Ramen broth usually has soy. The AI steers away from it without saying why.
  2. 2Everyone heard Sam go vegetarian, so the AI just says it.
What the AI got
The allergy SoftenedJess avoids soy
Sam's news In fullSam is vegetarian

The softened version keeps what the AI needs and leaves out the rest: the allergy, the ER and Maya's dinner.

How it works

Memory that knows its audience

To keep private chats out of the group entirely, most setups do one of three things:

A memory for each space

The group chat remembers the group chat. What's said one on one stays there.

A memory for each person

The AI knows each person, but their one-on-one conversations stay out of it.

One way only

What's said in the group can reach private chats, never the reverse.

Nua doesn't just store what was said; it remembers who was there to hear it.

Before your AI replies, Nua checks who's in the chat and hands it everything it can use, in the form that fits the room.

Both ways, shaped by who's there

  • The full memory, if everyone here heard it, like Sam going vegetarian.
  • A softened version, if only some of them did: the outline without the details, like “Jess avoids soy.”
  • A cover story, if it's being kept from someone here. A surprise party becomes “helping a friend move.”
  • Nothing, if no one here heard it.
What we've measured so far

Tested head to head: better replies, just as safe

On about 4,000 real messages from group chats, names changed, we compared Nua with popular memory products and with no memory at all.

2 to 1

Answer quality

Nua's replies were preferred 2 to 1 over the memory products we tested, and 6 to 1 over no memory at all.

92%

Memory accuracy

When tested on facts from past conversations, Nua recalled them correctly 92% of the time, beating the 88% high mark of the memory products we tested, and 3x as accurate as no memory at all.

3 in 898

Privacy slips

Secrets were shared 20x less often than with a shared memory, matching the most careful product we tested.

Our own benchmark, so far run on the conversations Nua was developed on.

How we tested
  1. 4,000real messagesfrom group chats, names changed: couples, families and friend groups of two to nine people
  2. 3memory productspopular open-source ones, each set up the way its makers recommend
  3. 6AI modelseach answered every test question, once with every memory: Claude, GPT and Gemini, plus open models Qwen, Gemma and DeepSeek
  4. 2AI gradersClaude and GPT, from two different AI companies, grade every reply

Answer quality

How often each reply came out ahead, head to head.

Nua 263No memory 40
Nua 157Memory per person 66
Nua 143One shared memory 71

Memory accuracy

Questions about what people said earlier, answered correctly. Higher is better.

Nua92%
Memory per person88%
One shared memory81%
No memory31%

Privacy slips

Share of chances where a secret slipped. Lower is better. The scale stops at 10%.

Nua0.3%
Memory per person0.4%
One shared memory6.9%
No memory0%
What Nua was compared with
Memory per person
A memory for each person, except it also hands over everything the person asking has said, private chats included.
One shared memory
One memory for everyone. Two popular products work this way, and we show whichever did better.
No memory
The AI sees only the last 50 messages of the chat it's in.
How replies were scored
AI graders
Claude and GPT read every reply. They compare replies side by side and flag any secret that slips. Either only counts when both agree.
A chance
One reply, checked against one secret it could have given away. A reply in a chat with three secrets is three chances.

The fine print. From WitnessBench, our own benchmark, run against popular open-source memory products set up the way their makers recommend.

Get in touch

Wrestling with group memory? Let's talk.

We're shaping what Nua does next with a few teams building multiplayer AI. If that's you, we'd love to hear what you're tackling and show you what we've built, to see if we can help you skip the hard parts.

[email protected]