Your clone's memory.
Your clone has two kinds of memory. They do different jobs, and knowing the difference tells you what to expect from it.
Dynamic memory: what follows you
Dynamic memory is the always-on layer. It is a living model of you: your projects, your decisions, your way of working. It follows you across sessions, so a new conversation starts already knowing the context instead of starting from zero.
You do not manage it. It grows as you work. When you tell your clone something that matters, it keeps it.
On hosted, the conversations that feed it are yours to take out. Ask your clone to export your turn history, for example "export my conversation history to a file", and it pulls every turn of every session you took part in, both your side and your clone's, oldest first, with your own sign-in. The export tells you when it has everything, and if your history changes while it runs, it says so and starts over rather than leaving a gap. The narrow key your clone uses to read context on every message cannot do this.
What your clone has concluded about you is yours to take out too. Ask "export what my clone believes about me to a file", and it pulls the cards it keeps about you, then every conclusion it ever stored about you, each with its date, not only the recent ones it shows while you work. The export tells you when it has everything, it only ever contains what was derived about you, and the same rule applies: your own sign-in or your account key, never the narrow per-message key.
Dormant memory: the deep archive
Dormant memory is the searchable history of everything: your past conversations, your notes, your imported ChatGPT and Claude history. It is not injected into every conversation. It sits there, indexed and dated, until your clone reaches for it.
That is what makes "the call from three weeks ago" recallable instead of gone. Last year and last week never blur into one, because everything keeps its real timestamp.
What gets remembered, and when
- During the conversation, your clone writes down what matters as you talk.
- When a session ends, a sweep stores the conversation into dormant memory so it stays searchable later.
- Before anything leaves your machine, credential-shaped strings (passwords, API keys, tokens) are redacted. Your secrets never enter memory, hosted or local.
Klyr only remembers what you give it inside Klyr. It never reads your machine or your files behind your back.
Searching the past
There is no search box to learn. Just ask your clone:
"What did we decide about pricing last month?"
"Find the conversation where I described the onboarding problem."
It searches its dormant memory and comes back with what it finds, dated. If it does not reach for the past on its own, tell it to look. "Search your memory" works.
You can also point a search at one store: "search only my Fathom meetings" or "look in my ChatGPT archive". Any store in your inventory is a valid target.
Seeing what memory holds
Run /klyr:status to see your memory inventory: every store — your notes, your conversation history, each import you brought in (Fathom meetings, a ChatGPT archive, ...) — with its document count. A real count, an explicit "empty but healthy", and an explicit "unknown" are three different states, and an unknown count is never shown as an empty memory. The inventory is scoped to your account on the server side, always: what you see is yours and only yours.
Correcting or removing a memory
Memory is not a black box. If your clone remembers something wrong, tell it plainly: "That is outdated, I no longer work with X. Forget it." It updates or removes the memory. You can also ask it to show you what it believes about any topic, then correct the record line by line.
Bulk mistakes are removable too. If a whole import went in wrong — the wrong export, a batch you never meant to keep — ask your clone to delete that batch, or a single session, from dormant memory. It first shows you exactly how many entries match, and erases only after you confirm. Only your own memory is ever touched.
Batches can also be renamed without losing anything. If a slice of your history is filed under the wrong project label — an old machine path, a mislabeled import — ask your clone to relabel it. It shows you how many entries match first, then rewrites just the label after you confirm: the memories themselves, and their searchability, stay exactly as they were.
Your identity files (who your clone is, who you are) are plain text on your machine. You can open and edit them yourself anytime.
Taking your dormant memory with you
On hosted, your dormant memory (your notes index, your conversation history, each import) can be read back out with your own sign-in, no request to us needed. Ask your clone to export a store, for example "export my Fathom store to a file", and it pulls the entries page by page into a file on your machine. Each entry comes out in the same shape a local install writes, so the file is usable whichever way you run Klyr.
An export runs one store at a time, or every store at once, and it tells you when it has everything: an export that stopped halfway never looks finished. It only ever returns your own entries. The narrow key your clone uses to read context on every message cannot do this; only your own sign-in or your standing account key can.
This covers dormant memory. The dynamic layer (the living model of you) has its own two exports, your conversation history and what your clone concluded from it, described above.
Hosted or self-hosted
Both memory layers work the same either way; only where they run differs. Hosted: your memory runs in your own private space on our servers, one space per person, never pooled, never used for training. You sign in once with /klyr:login and your clone follows you to any machine. Local: everything runs on your own computer in your own Docker stack. Free, nothing leaves the machine, tied to that one computer. You chose between them during /klyr:init, and you can change your mind later.