Memory, explained: Memory Spaces, Fold, Engram and sleep

Halfway through an unrelated conversation you mention that Halton pay on 45 days rather than 30. Three weeks later, in a fresh chat, you ask what terms Halton are on and get the right answer.

None of that happened inside the AI. Four things did the work: somewhere to write the note, a way to find it again from a question phrased nothing like it, something that keeps the collection in order, and a background pass that does the keeping.

The model does not remember anything

An AI model reads the conversation in front of it and nothing else. It has no idea what you said last Tuesday unless those words are on screen with it now. That is true of every model from every provider, and a bigger one does not change it.

Everything that survives between conversations is written down by Player One and read back by Player One. When an agent appears to remember you, it searched before answering and found a note it wrote earlier. That also explains the failure you will occasionally see: if the note was never written, or the search did not surface it, the agent carries on cheerfully without it.

Conversations themselves are a separate business, and a long one gets summarised as it grows. Conversations covers that.

Memory Spaces: somewhere to write things down

A Memory Space is a named collection of notes with a stated purpose and audience. Every agent has one that follows it everywhere, and you can create shared ones for anything worth keeping in a single place: your suppliers, your policies, everything known about a major customer.

They are ordinary files, and you can read every one. Open a Space from Manage > Memories (/manage/memories) and its Files tab shows notes waiting to be filed, source documents kept whole, settled write-ups by subject, and a dated history. You can edit or delete any of it by hand. Memory you cannot inspect is memory you cannot trust.

The other tabs carry the Space's own instructions: what it is for, when recalling from it helps, what a useful note looks like. Access sets the audience, a real boundary rather than a label: internal means you and your agents, public means a support bot or a customer may read it, so nothing confidential belongs there.

Fold: finding it again by meaning

Writing notes down is the easy half; finding the right one later is the hard half, and that is Fold, which indexes a Space by meaning rather than by exact words. Ask about "the supplier who was late in March" and it surfaces a note reading "Kesteven missed the delivery window on the 14th", although the two share no significant word. You never have to remember how you phrased it.

This is proto-holographic recall: what it finds is not one sentence but the neighbourhood around it, the decisions, files and history connected to that note, weighted by what you actually use. Memories you keep recalling stay strong; ones nobody has needed for months fade towards the background and never disappear. Fold describes it in full.

It works this way because a model can only use what fits in the conversation with it, and that room is finite and paid for by the word. Retrieval has to be precise and small: a handful of genuinely relevant files, not everything that ever mentioned Kesteven. Keyword search fails in both directions, missing the note that used other words and burying the good answer under forty weak ones.

Each Space has a Fold tab (/manage/memories/<slug>/fold), where you can run that search yourself, see what is indexed, see any file it skipped and why, and check what it has cost. Admin > Fold (/admin/fold) shows every index at once, with pause and re-index controls. A Space that is not indexed cannot be recalled at all; new ones are indexed automatically, and the page says when one is not.

Engram: the librarian

Fold indexes what it is given, and what it is given is messy. Notes are written in the middle of doing something else, so they arrive in the order things happened rather than by subject: three saying roughly the same thing, one covering four unrelated topics, a few already stale. Ask about that supplier and six overlapping notes come back, filling the room that should have held the answer. Precision comes from the collection being well shaped, not from the search being clever.

Engram is the agent that shapes it, one Space at a time. It reads the notes that are waiting, works out what each is genuinely about, and folds them into the existing write-up on that subject rather than starting another file. A note covering three things is split across the three places it belongs. Anything that mattered on a date goes into the Space's dated history, with a short recap for each month and completed year, so the sequence survives as well as the conclusions. Related files are linked, so the Space reads as a small wiki, and its Read and Write pages stay accurate.

Each Space also carries a short recall index in four tiers, from the core knowledge that is never evicted down to what arrived this week, and Engram alone moves entries between the tiers as their importance and use change. It keeps everything, files what it cannot place in an unsorted area it comes back to, and repairs its own structure as it goes. Engram describes it in full.

What it never does is throw anything away. Something it cannot confidently place stays searchable rather than being tidied out of existence, because a memory that looks irrelevant today may be the one you need next year.

Sleep: when the librarian works

All of that is model work, and doing it in your conversation would slow every reply and spend your tokens on filing, so it happens out of the way in a background pass called sleep.

Sleep is not a nightly batch. A coordinator watches every Space and starts a pass when there is enough to justify one: by default around fifteen notes waiting, or a couple of days passing with anything waiting at all. One Space is never started twice within the hour, and only a couple run at once, so maintenance never crowds out your work. Both thresholds are yours to change on the Space's Sleep tab, and you can turn sleep off for a Space you curate by hand.

Admin > Engram (/admin/engram) is where you watch it: what is running, what is queued, each Space's schedule, and a History tab recording how every past pass ended. It is cautious by design: if passes start failing it stops and waits for you.

So a note is saved and searchable the moment you ask for it, but it will not be tidy for a while. That is normal; asking twice does not help.

Which Spaces to have

A Space is defined by what it remembers and who may read it, not by a shape you pick in advance. The recognisable ones: a personal notebook of contacts and preferences, an internal wiki of runbooks and decisions, a public knowledge base your support bot answers from, a searchable archive of whole documents.

You are never asked to design folders or pick a category scheme. You say what you want to remember and who may read it, and the shape emerges from the material that arrives, because guessing a structure before you have content produces one that fits nothing. Build a knowledge base walks through filling one.

What this means for how you work

Say "remember this" when something matters. It is a normal sentence, not a command, the note appears on the Memories page at once, and then you let it sleep.

Skim the Memories page occasionally, the way you would a filing cabinet: correcting a wrong memory takes seconds and saves an argument later. Keep contexts apart with projects, so a shared Space is only in play where it belongs, as Organising with projects sets out.

A correction to how something should be done usually belongs in a skill rather than a memory, because a skill is loaded when the work happens instead of being searched for. Teaching it explains where each lesson goes.

Where memory sits beside your files is Workspaces and Memory Spaces. The short description is Memory, and the same ground told through examples is It remembers.