Feature
Niyra remembers
You have a long conversation with most AI assistants and then start fresh tomorrow. Niyra doesn't work that way.
How it works
Niyra's memory has three layers:
1. Session search. Every conversation Niyra has had is indexed with full-text and semantic search. When you say "how did we handle that last quarter?", she finds the actual conversation and quotes the relevant exchange.
2. Vector memory. When you tell Niyra something worth keeping, a preference, a fact, a date, a person, she distills it into a structured memory and stores it with an embedding. Categories: preference, fact, person, place, date, work, health, finance, general.
3. Records. For structured data with expiry, vehicles, policies, subscriptions, holdings, Niyra uses typed records with reminders. (See Records.)
What she remembers
- Preferences. "I prefer 9am meetings, not before."
- People. Names, roles, kids, anniversaries, last conversation.
- Dates. Birthdays, renewals, insurance expiry, exam dates.
- Decisions. Past choices you can refer back to.
- Working style. How you write, what you skip, what you delegate.
What she forgets (on purpose)
- One-off chitchat with no actionable signal.
- Anything you say "don't remember this" about.
- Anything you delete from the Memories panel.
Why it matters
Most assistants either give you no memory (start fresh every conversation) or give you "infinite context" (remember literally everything, including the noise). Niyra picks what's worth keeping, the same way a good chief of staff does. You don't have to manage her memory; you can if you want to.
Measured, not claimed
We ran this memory system against LongMemEval, the hardest public benchmark for long-term conversational memory: 500 questions across 246,738 conversation turns. The production retrieval stack, zero benchmark tuning, scored 98.2% recall@10 and 92.0% recall@5, including 100% on knowledge-update questions (the "this fact changed, do you know the current truth?" category). We measured at turn level, which is roughly ten times finer than the session-level numbers most memory systems publish, and we published every caveat alongside the numbers. Read the full benchmark write-up.
Try it
The fastest way to see the difference: tell Niyra something on Monday, ask her on Friday. She'll answer like she heard you the first time.
FAQ
- Is memory the same as chat history?
- No. Chat history is a transcript. Memory is structured facts Niyra distilled from your conversations, "user prefers morning meetings", "user's wife's name is Priya", "user's RC expires Aug 2027". Memory survives across sessions and is retrieved semantically when relevant.
- Can I see and edit my memories?
- Yes. The Memories panel lists every memory with what it says, when Niyra learned it, and last accessed. Edit any, pin important ones, delete anything you don't want her remembering.
- Does memory work on WhatsApp / Telegram?
- Yes. Memory is account-level, not channel-level. Tell Niyra something on WhatsApp, she'll remember it next time you chat on web.
- How does retrieval work?
- Embeddings (text-embedding-3-small) over memory content. On every message, Niyra runs a semantic search and pulls the top relevant memories into her context, but only the ones she actually needs, not all of them.
- How good is the memory, really?
- Measured, not claimed. On the full LongMemEval benchmark (500 questions, 246,738 conversation turns), the production retrieval stack scored 98.2% recall@10 with zero benchmark tuning, including 100% on knowledge-update questions. The full write-up publishes every number and every caveat.