Memory by meaning, not keywords
Every Ermes agent ships with a real vector database — the same technology behind modern AI search — as its long-term memory, and a visual Memory Atlas so you can actually SEE what it knows. Here's what that means, in plain words.
The Memory Atlas — a live demo on example memories. Every dot is a memory, related things sit together, recalls fire like synapses. Your own map is private to you.
Meaning becomes a place
A vector database stores every memory as a point in a huge “meaning space.” Things that mean similar things sit close together — “faktura za serwer” lands next to “hosting invoice,” even though they share zero words and aren't even the same language. When you ask your agent something, it doesn't search for matching words — it goes to the right neighborhood of meaning and reads what lives there.
Ask the way you actually think
Human memory doesn't work in keywords, and neither do your questions. “That invoice thing from a few weeks ago,” “what did the client say about the deadline,” “the restaurant Marta recommended” — none of these contain the words that were originally used, and all of them just work. Cross-language, cross-source: a WhatsApp message, an email, and something you told your agent directly all live in the same meaning space, so one question finds all three.
Chat history is not memory
Most AI tools keep a scrolling transcript, and a small “memory” of a few dozen notes the model decided to keep. That's a sticky note, not a memory. Ermes stores everything you choose to connect — tens of thousands of memories on the busiest agents — and recalls across all of it in under a second. And because it's a real database on a real server, it's yours: inspectable, searchable, erasable.
Private by architecture, not by promise
The entire memory — the vector database included — runs on a dedicated server that belongs to you. Nothing is pooled, nothing is used to train anyone's models, and when you tell your agent to forget a source, the memories and everything derived from them are actually deleted. That's the difference between a privacy policy and a privacy architecture.
The Memory Atlas — see what your agent knows
Because vector memory places things by meaning, it can be drawn (that's the orb above). The Memory Atlas renders your agent's entire memory as a living map — every memory a star placed by meaning, people and projects as their own hubs, related things clustered no matter where they came from. When your agent recalls something you watch the memories fire between each other like synapses; when it learns something new, a star appears and wires itself in. Search by idea and the map lights up; click any star and it tells you why it's there and which conversations used it. Ask your agent “show me my memory map” and it sends you a private, full-screen link — no login needed.
A second brain — living, cross-linked notes
Raw memory is the foundation; the second brain is what your agent builds on it. Every night it distills your memories into living pages — one per person, project, and topic — and rewrites them as things change instead of piling up. The pages cross-link like a wiki (a note about a deal links to the people and projects it touches), every page cites the exact memories it came from, and when you tell your agent to forget a source, everything derived from it is erased too. It's the pattern Andrej Karpathy has described publicly — keep everything, distil the important part into a maintained wiki — built into your agent. See how the second brain works →