The whole idea, in plain English

The problem

An AI chatbot is a very smart person with no memory. Every time you open a new chat, it is like meeting them for the first time. You told it on Monday that you are vegetarian. On Thursday it suggests a steak. Nothing you say survives the end of a conversation.

The first fix everyone tries is to paste the whole past conversation into every new chat. That is like reading your entire diary out loud to a friend before asking them one question. It works for a week. After a year the diary is longer than the friend can listen to, and you pay for every page every time.

The fix: a card file

Instead of the whole diary, keep a small box of index cards. After each chat, an assistant writes down only the things worth keeping: "Maya is vegetarian." "Maya's partner Sam has a birthday on June 18." Small talk and jokes do not get a card.

Before the next chat, you do not hand over the whole box. You pull the three cards that match the question and hand over only those. Ask about dinner and you get the vegetarian card. Ask about a gift and you get the Sam card. Three cards today, three cards in ten years. The cost never grows.

That box is the memory layer. Everything else is about keeping the box small, true, and easy to search.

Two people run the box

The secretary writes cards. It reads the chat once, at the end, and writes full sentences with names and dates: "Maya moved from London to Paris in March 2026", not "she moved". It works in the background, so you never wait for it. It is allowed to write a new card and to say "this card replaces that older one". It is never allowed to tear a card up. This one rule turned out to be worth twenty points on the public tests. If you tear up "lives in London" when you write "lives in Paris", you can never again answer "where did I use to live?"

The librarian finds cards. It searches three ways at once: by meaning (a question about dinner finds the vegetarian card even though the word "vegetarian" is not in the question), by exact words (a question about "Poppy" finds the card that says Poppy), and by names (a question about Sam finds every card that mentions Sam). Then it prefers cards that are important and cards used recently, and it hands over the top few with their dates written on them.

Every card has two dates

A card is not a cell in a spreadsheet. It is more like a commit in git. It says when the fact was true in the world and when we learned it. When Maya moves, the old card is not destroyed. It is closed, with a note that says "replaced by card 7 in March 2026". Ask "where does Maya live?" and you get the open card. Ask "where did she live before?" and you get the closed one. Every card also says which conversation it came from, so a wrong card can be traced and so that facts from the user's own mouth can be trusted more than facts copied from a random web page.

Cards have three colours

Facts ("Maya is vegetarian") stay until something replaces them. Events ("Maya planned a Lisbon trip on June 21") are dated and fade after about three months if nobody asks about them, like old diary entries. Rules ("answer short and direct") never fade and are always handed over, because they are tiny and they shape every answer. The colour decides how long the card lives.

Why forgetting matters

A box that never throws anything away is a hoarder's garage. Five cards say where Maya lives, four are stale, and the librarian pulls a stale one first. So: never write a duplicate card, close old cards instead of stacking new ones next to them, let unused event cards fade, and keep one switch that empties a whole person's box when they ask. Forgetting is not losing information. It is what keeps the top three cards correct.

Where the ideas come from

Your class notes gave the shape: secretary, librarian, box, three colours, forgetting, and one rule that every search is limited to one user's box. Mem0, the popular open-source library, taught us the secretary should write rich sentences, never delete, and the librarian should search three ways at once. HydraDB and Zep taught us the two dates on every card and the git-like "replaced by" pointer. The benchmark papers told us what the tests punish: facts that changed, questions that need two cards, and questions about time.

How we will know it works

Three public exams exist. LoCoMo asks questions about very long conversations. LongMemEval asks about facts that changed. BEAM asks about conversations of a million words or more. A score only means something next to how many cards were handed over, because handing over two hundred cards is easy and expensive. We will publish both numbers together.

What we are building

One small Python library. One SQLite file per user bucket holds the cards, the original chat lines they came from, a name index, and a word index. A few small files of code: one for the secretary, one for the librarian, one for the box, one that turns "last summer" into dates. Any chat model behind a standard API. Later, when there are a million users, we shrink the numbers on each card nine times with a trick called TurboQuant, and we never have to retrain anything to do it.

That is the whole system: a secretary who never tears up cards, a box where every card has two dates, and a librarian with three ways to search.