The agent that confidently did the wrong thing
Why we are building Lore, in the open.
It started with a demo that worked perfectly and was completely wrong.
We had wired up an agent to handle support tickets. It was fast, it was articulate, it sounded exactly like one of us. Then it approved a refund it never should have, citing a policy that had not been true for six months.
The model was not the problem. It was as capable as it had ever been. It just did not know how our company actually worked. The real refund rule lived in a Slack thread from last spring, a follow-up email from finance, and a Notion doc that three people swore was out of date. The agent had no way to know which one was current, so it guessed. Confidently.
The thing nobody was building
Everyone we knew was making agents smarter. Better models, longer context, more tools. Almost nobody was making them know the one thing that makes a teammate useful: how this specific company does this specific thing.
There are memory tools for this, sort of. Mem0, Zep, Letta. But they store notes the agent reads and improvises from. We did not want our agent to improvise our refund policy. We wanted it to follow the policy, with the receipt attached.
What we actually built
Lore reads the places your decisions actually live (Slack, Gmail, Notion, even our own Claude Code sessions) and writes them into SKILL.md files, Anthropic's open Agent Skills format that any agent can load and run. Every line cites the exact source that set the rule. When the rule changes, the skill re-emits and the citation moves to the new source.
Not a wiki for an agent to read. Process for an agent to run.
Memory remembers. Lore acts.
We run Lore on Lore
The most honest test we have is that our own company runs on it. Three cited skills so far, written from nothing but our real work: refund-handling from 47 real observations, incident-response from 41, pricing-exceptions from 38. All auto-written, all cited back to the conversations and decisions that created them. We would rather show three you can audit than thirty you cannot. The agent that drafts our marketing reads that same brain. So does this account.
Where we are
We are early. In YC-review build, hand-onboarding the first design partners, free during the first cohort. It is rough in the places early things are rough. But the core bet feels more right every week: your agents do not need to be smarter. They need to know how you actually work.
If your agents are running on tribal knowledge buried in Slack, we would love to talk. That is the whole reason we are building in the open.