If you use ChatGPT or a cloud assistant, you know the problem: with every new conversation, it starts from scratch. It has forgotten who you are, what you explained yesterday, the project you have been working on for three weeks. This isn't a glitch, it's how they work by default: these assistants have no memory from one session to the next. The only real fix is an assistant designed to genuinely persist a memory (your past exchanges, your documents, a context that builds over time), on infrastructure you control. That's the design choice behind Gungnir.
Two reasons stack up. The first is technical: a large language model has no memory of its own. It only sees what you hand it in its context window on each request. Outside that window, nothing persists, unless the service decides to store your history and feed it back in.
The second is more awkward. To remember you, a cloud service has to keep your data on its servers. The more it knows you, the more personal information it accumulates outside your control. Memory and confidentiality pull in opposite directions, and most providers settle it by keeping the bare minimum, or by keeping everything on their side. Forgetting is often the easy answer for them, not for you.
For an assistant to truly remember, you need more than a chat history. Gungnir stacks three layers, and each one stays inspectable.
1. Conversation memory. Every exchange is saved and reloaded in the next session. The conversation also follows the person from one channel to another: a message received on Telegram picks up the ongoing thread instead of replying into the void.
2. Your documents. You build a knowledge base from your files. The assistant goes looking for the information when your request calls for it, instead of pretending to know or making up an answer.
3. A context that consolidates over time. In the background, the assistant turns its short-lived notes into durable memories, rereads itself to spot its own contradictions, and measures the gap between what it expects and what actually happens. This is the layer we call "consciousness", and the word deserves a clarification.
Let's be clear, because the term is overused. Gungnir does not claim to be conscious in the strong sense, and our own documentation states it plainly: we describe a measurable cognitive architecture, not phenomenal consciousness. Words like "need" or "mood" that appear in the system are shorthand for deterministic, auditable mechanisms, not claims to a lived experience.
In practice, this layer does three useful things. It consolidates memory, so the important details don't get lost in the flow. It critiques itself, with a module that actively looks for its own inconsistencies and contradictions. And it measures its reliability, by comparing what it predicts to what actually happens. Nothing magical in there: loops, measurements, and files you can open.
The decisive difference isn't only that Gungnir remembers, it's where those memories live. They are stored on your infrastructure, in your database, with a dedicated panel where you can review the memories it keeps and erase them if you decide to. Where a cloud assistant asks you to hand over your memory so it can know you, Gungnir lets you keep it.
It's the same principle that has the assistant running on your own server: what it learns about you never goes anywhere else. And if "who can read my data?" is what holds you back, we covered it in detail in the piece on cloud assistants and the GDPR.
Yes. Every exchange is saved and reloaded in the next conversation, and the thread even follows the person from one channel to another: a message on Telegram picks up its discussion. You start from context, not from zero.
Yes. You build a knowledge base from your files, and the assistant draws on it when your request warrants. It searches your own sources instead of guessing.
No, and we don't claim it. It's a measurable cognitive architecture (a memory that consolidates, self-critique, a measure of its own reliability), not a lived experience. Words like "mood" or "need" are shorthand for deterministic mechanisms, not feelings.
On your infrastructure, in your database, with the embeddings going to the vector engine you configure. A dedicated panel lets you review the memories it keeps and erase them. Nothing is kept on a third-party server.
