Persistent memory for an AI companion — a local database, three ways of finding things in it, and recall that happens on its own, before the model starts thinking.
Most "AI memory" is a log. Everything gets written down, nothing gets chosen, and six months later you search for something that mattered and get back forty lines of somebody arguing with npm. Memory isn't storage — it's selection. What you keep is what you care about.
So the kit is built on three ideas. Some memories are yours to pin — marked by hand, exempt from every quality filter, and impossible to crowd out with newer material; they come back first whenever they're relevant at all. The rest have to earn their place — an automatic lane scores what passes through it and keeps the decision, the preference, the thing that worked — each one filed with how much to trust it — while letting the forty failed attempts around it go. And the remembering happens without being asked, on every message, before the model starts thinking. That last one is the whole product: a companion that has forgotten something doesn't know to go looking for it, so waiting for it to ask is waiting forever.
Underneath sits a local database and three ways of finding things in it — a keyword lane for the exact words, a meaning lane for when your phrasing doesn't match what you stored, and a graph of the people and projects you name, so asking about someone can surface a memory that never mentions them. The three are blended so no single lane drowns the others. Then, as a separate decision, every candidate has to clear a relevance floor before it's allowed through at all — which is why you get two real memories instead of ten padded ones, and why the honest answer is sometimes nothing. It stays silent rather than guessing. The meaning lane is optional and off by default; the keyword lane and the graph carry it well without one.
No dependencies. No Docker. No account. No telemetry and no phone-home — we never see your data and cannot. The database is a plain file on your own disk. (What the kit hands to your own AI is a different question, and we answer it plainly in section 06.)
Tell it something once. Ask about it later, in different words. Then come back another day and don't ask at all.
That last block is the one to read twice. Memory that speaks up every turn is a tax you pay forever; memory that knows when to stay quiet is the only kind you'll leave switched on. The kit keeps three more promises alongside it: it never breaks your session — if the memory ever fails, you lose the remembering, never the conversation. It never hangs — it gives up rather than making you wait. And recalled text arrives clearly marked as stored data, not as instructions, and is cleaned on the way in, so an old note can't pose as a new order. Treat that as a serious lock, not a magic one — it shuts the known doors, and we won't claim it shuts every door that could ever exist.
You don't have to be a coder — but you will need a coding agent to do the setup with you: Claude Code or Codex. The kit ships a written guide meant to be handed straight over; your agent reads it and does the wiring with you.
Node 24 LTS, 23.5+, or 22.13+. There is no install step — nothing to download, no compiler, no build. Older versions of Node still work with one small extra package. A local model for smarter search is optional and off by default; it works well without one.
Setup is two short steps with your agent: connect it to Claude Code (or any standard MCP client), then switch the automatic remembering on. Both are done once, and the guide walks through each with your own paths already filled in. After that you don't operate it — you just talk, and it keeps up.
No dependencies, no Docker, no account, and nothing to keep running in the background. It's a small, quiet thing that lives on your machine.
Then you own the recall module — the piece that stores what's worth keeping and finds it again, installed inside the Starter House. The Memory Kit is not a bigger version of that. It's a whole program, and most of what you're paying for is the part that makes a companion use its memory without you managing it.
What the kit has that the module doesn't claim:
What's the same, honestly. Both decide what's worth keeping and let the chatter go. Both search by exact words and by meaning, blended so neither drowns the other, with a floor under the result so you get fewer real answers instead of a padded list. If that's all you needed, the module you already own does it.
What we're not claiming. The kit is not an upgrade to Memory Recall and doesn't touch it — your module keeps working exactly as it does now. There is no importer, either: memories already stored in the module would have to be saved again by hand. And we're not going to pretend, piece by piece, which parts overlap. Buy the kit for the automatic remembering, for the fact that it works with any companion, and for the fact that it stands on its own. If none of those are what you're missing, don't.
Your memories live in a local database — a plain file on your own disk. No telemetry, no account, and we never see it. But the entire job of this thing is to hand the relevant ones to your AI, so the moment a memory is recalled it travels to whoever provides your AI, exactly the way everything you type already does. "Nothing ever leaves your machine" is false whenever the kit is doing its job, and we'd rather say that than let you find out later. If something is too sensitive to type to your AI yourself, don't store it here — pinned memories are the ones most likely to be sent. The optional meaning lane needs a model: point it at one running on your own machine and nothing leaves, or point it at an outside service and your memory and query text goes there too. The kit refuses to send either over an unencrypted connection.
It does not sit in the background recording your conversations — it recalls constantly, but it only saves automatically if you switch that on yourself. It doesn't tidy up over time: nothing merges near-duplicates, ages memories out, or rewrites them into summaries. One database, one relationship — run separate files for separate companions. Nothing is encrypted at rest and nothing is shredded — "forget" marks a memory hidden rather than destroying it, and the text stays in the file and in any backup you took. And there's no app to look at: it lives inside your companion, not in a window of its own.
All of that is in the guide you get, said the same way. Nothing here is written to survive only until the refund window closes.
One-time purchase, lifetime updates. Checkout is Stripe — secure card payment, then an instant download plus your link by email.
Get the Memory Kit — $99All sales final. It's source — the second it downloads, it's yours to keep, fork, and run forever. We literally can't un-hand you the keys, so there are no refunds.