Dreaming teaches agents to learn from sessions. What about everything you wrote before it?

Anthropic shipped Dreaming in May, a way for Claude agents to review their own sessions while idle and write memory notes that improve future runs. It's a real capability, and it's also a useful example of a limit that's still true two months later: Dreaming is scoped to sessions, to what an agent did during its own runtime.

That scope is specific to agents. But for us as humans, many of us carry a much larger body of material accumulated over years, long before this recent wave of AI, that stays untouched no matter what any agent remembers about its own sessions.

The problem Dreaming doesn't solve

I've been taking notes for almost 20 years across a career in voice, networking, and cybersecurity, through client meetings, research, strategy work, and market interviews. I went independent last year, and it was in the thick of that transition that I hit a point where I was spending more time searching for old thinking than generating new thinking, which is what pushed me to build a personal tool that solved it for myself. This year, I turned that proof of concept into Elicana: a self-hosted tool that compiles years of accumulated documents into a normalized, queryable knowledge base, so you, or even your AI agent, can find something you wrote in 2019 just as reliably as something you wrote last week. It runs entirely on your machine, so nothing is ever uploaded to build it.

I'd captured copious amounts of notes over those years, but identifying patterns through them meant running into a few different challenges. One of them was vocabulary drift: in 2019 I called something "client onboarding friction," by 2021 the same underlying problem had become "deal velocity issues," and by 2023 it was "sales cycle drag." Three notes, one insight, three different labels, and search returns whichever phrase I happen to type, on a good day.

This isn't really a search bug so much as a data quality problem, and it's structural: a corpus built over years accumulates inconsistent terminology simply because the person writing it changes, new clients, new industry language, new reading along the way. Twenty years of notes look inconsistent by default, not by any failure on my part.

Pattern recognition across engagements is what I actually charge for as a consultant, so when I can't reliably surface my own prior work, I end up wasting time hunting for the pattern in my own corpus instead of using that time productively. It's time I can't bill the client for.

Why the standard tools don't close this gap

Cloud AI knowledge tools ask you to point a folder at their servers and index whatever's there. Some require careful curation, keeping the sensitive client notes out and managing what's in scope, while others just ingest everything. Either way, they index what you wrote, not what you meant, and none of them resolve the vocabulary gap, because doing that requires understanding across the whole corpus rather than matching a query against a single chunk.

Dreaming is a good instinct aimed at a narrower problem than this one. It teaches an agent to learn from its own sessions, but it does nothing for the twenty years of sessions that happened before the agent ever existed, and every knowledge worker with more than a couple of years of accumulated notes has exactly that gap, whether or not they've ever heard of Dreaming.

What Elicana does about it

Before you ever search your archive, Elicana reads through it once and works out which notes are really talking about the same thing, even when you used different words for it at different times. "Client onboarding friction," "deal velocity issues," and "sales cycle drag" get connected as the same underlying concept, so a query for any one of them finds all three. It runs locally, and your archive never leaves your machine.