They promised me Jarvis… they gave me Drew Barrymore in 50 First Dates. I will set off all your BS trip wires, but you should still try my Hermes alternative anyway. I make no money at all from it and it’s S tier good.

Let me get the dodgy stuff out first. I'm not a developer. I can't even type a function that sums two numbers. I'm a simple dude from Rome (Italy) who's decent at systems thinking, and Claude wrote the code. There are thousands of delulu people on Reddit with my same lack of skills confidently saying they built something amazing. They didn't. Their slop is painful and follows the same security practices as a heroin addict who forgot to buy new needles. I built this because I was fed up with agents forgetting stuff. Even the best memory systems rely mostly on retrieval. That's fine if you don't expect the agent to connect dots on its own. But, as Donald Rumsfeld would say, the agent "doesn't know what it doesn't know". Which is doubly troubling. Both for agent performance and for the fact that smart people used to be Secretary of Defense. To know what you don't know, you need to know something that hints at what you don't know. If that sentence didn't 404 your grey matter, you get the architecture. Instead of a .md file with facts appended on the go, Briglia spends up to 50k context tokens on a curated chronological history of its conversation with you. Think about it. A full day of work with an agent can burn millions of tokens, but what you and the agent actually said to each other could be just 10k. Those messages are already an amazing distillation of everything you worked on. Every decision, every preference, every "no, not like that" is in there. Compact those 10k into a dense 1k summary and you've kept basically the whole day. Briglia orders these summaries chronologically and keeps curating them. Recent weeks stay detailed and older months get compressed, but the model always sees the full picture, up to a year back. On top of that, a fact-archival pass catches what gets diluted over time, and it can grep the exact old messages when it needs a detail. There are no sessions. It's one conversation, for months. The result is a slightly bigger context, but way fewer tool calls to reconstruct what the hell you're talking about. Six months ago that would have meant context rot. Today's models handle 100k-250k just fine. I have “biased” receipts of how good it is, it’s not just vibes. On my preliminary long-conversation memory benchmark it remembered 43 of 48 old facts on a month long conversation. Hermes + Hindsight got 39, Codex got 6. Method and data are explained on a link in the comments (yes, it's my own benchmark so it’s probably worth just slightly more than toilet paper to you - also, It’s a very preliminary version of the benchmark). Install command and repo in the comments. You need an OpenRouter API key and a Telegram bot token. That's it. Fair warning, it's a real agent with full access to your machine (shell, files, web, email). Run it on a computer you trust it with.