Knowledge & Memory - RAG Knowledge Bases + Self-Updating AI Memory | ToRun

An AI that actually remembers - and keeps it accurate

Most AI forgets you the moment a chat ends. The next conversation starts from zero, so you re-explain who you are, what you're working on, and what you already told it last week. And when an assistant does remember, it usually remembers wrong - repeating a fact from three months ago as if it were true today.

ToRun fixes both problems. Knowledge bases let you bring your own documents so answers are grounded in your material, not the model's guesswork. Long-term memory lets the assistant hold on to facts about you - but with one crucial difference: those facts are designed to stay fresh.

Open my memory · Try it in chat

Knowledge bases: answers grounded in your own material

Upload your documents into a knowledge base and the assistant retrieves the relevant passages before it answers. This is retrieval-augmented generation (RAG) - instead of hoping the model happened to learn something, the answer is built from text you actually provided.

Your documents are split into searchable chunks and indexed by meaning, so a question phrased differently from the source still finds the right passage.
Retrieval blends semantic search (understanding intent, via embeddings) with keyword search (catching exact terms), so names, codes, and rare phrases don't slip through the cracks.
The strongest results can be reranked by a cross-encoder before they reach the model, so the most relevant passage leads - not just the first match.
Grounded answers mean fewer hallucinations: the assistant works from your material instead of inventing plausible-sounding fiction, and it can show which passages it drew from.

Use it for a product manual, a policy handbook, a research library, a contract set, or your own notes - and get answers that cite what they're built on.

How retrieval actually works

When you ask a question, ToRun turns it into a vector - a mathematical fingerprint of its meaning - and compares it against the vectors of every chunk in your knowledge base. The closest matches by meaning come back, are blended with any exact keyword hits, optionally reranked for relevance, and handed to the model as grounding. The model answers from that retrieved context rather than from memory alone. The result: rephrase the question, use a synonym, or ask in another language, and the right passage still surfaces - because retrieval is matching concepts, not just strings.

Memory that ages gracefully

Here's the part nobody else gets right. A fact isn't true forever, and not all facts go stale at the same speed. Your birthday never changes. Your job might change in a year or two. Your current project could change next month. ToRun treats each fact according to how fast it actually changes.

Invariant facts (birthday, name) never expire - and anything derived from them refreshes on its own. Tell it your birthday once, and your age is correct every single day after, with no "I'm 38" left over from last year.
Slow-changing facts (your city, your role) carry a gentle freshness clock, so the assistant knows when it's time to double-check.
Fast-changing facts (this week's focus, a current figure) fade faster and are flagged sooner.

When a fact is getting old, the assistant doesn't pretend it's current - it tells you "this was true as of three months ago, you may want to confirm." And it never silently throws old facts away: outdated information is ranked down and labeled, never deleted out from under you, so you can re-confirm or update it on your terms.

Find what you remember, in any language

Memory search is built for a multilingual world. Look up a remembered fact and ToRun matches it regardless of case or accent marks, and without one language's rules breaking another's - so searching "isim" finds "İsim," and an English query doesn't trip over a Turkish character set. Your memory is as searchable as it is accurate.

You stay in control

Memory should be something you can see and edit - never a black box. A dedicated memory view lets you read everything the assistant remembers about you, correct a fact, set how fast it should go stale, or remove it entirely. Everything is stored privately and tied to your account - it's your memory, and you own it.

Knowledge and memory work everywhere the assistant does - in chat, in research, behind your AI personas, and inside the in-app assistant - so a single, accurate picture of you and your material follows you across the whole platform.

What you get

Your own knowledge bases bi-folder2-open - upload manuals, handbooks, research, or notes and get answers built from your material, not the model's guesswork.
Search by meaning and by keyword bi-search-heart - retrieval blends semantic understanding with exact-term matching and optional reranking, so rephrased questions and rare names both find the right passage.
Facts that refresh themselves bi-calendar-check - store your birthday once and your age is correct every day after; derived facts update automatically.
Staleness, flagged not hidden bi-hourglass-split - when a fact gets old the assistant says "true as of a while ago - confirm?" instead of stating it as today's truth. Nothing is silently dropped.
A memory you can see and edit bi-sliders - read everything the AI remembers, fix a fact, set how fast it should go stale, or remove it.
Private and yours, everywhere bi-shield-lock - knowledge and memory are stored privately, tied to your account, and follow you across chat, research, personas, and the in-app assistant.