A phoenix called Scatterbrained

There is a particular grief in taking down your own work.

A few days ago I retired a project I'd named Engram — a personal knowledge graph I'd been building for months, the thing I opened every working session. After I published it I found ly-wang19/engram: an earlier, paper-backed (arXiv:2606.09900) bi-temporal memory engine for LLM agents. Same name, same bi-temporal idea of keeping a fact's history instead of overwriting it. They were first, and they did the rigorous, benchmarked work to earn it. So I retired the name and gave the thunder back.

I expected that to be the end of it. Instead it was the most useful thing that happened to the project.

What the name was hiding

"Engram" had quietly dragged my project into someone else's lane. The word wants to be said next to memory engine and AI-memory infrastructure — benchmarks, recall curves, a thing you embed in an agent. And every time I described mine that way, it felt a half-size too big, like a borrowed coat.

Losing the name forced the question I'd been avoiding: what is this thing, actually?

It isn't a retrieval engine. I never benchmarked recall, because that was never the point. What I'd actually built was quieter and more personal — a place where everything I care about (projects, decisions, the people who shaped them, the sources that back them) lives as one connected graph that an agent tends across sessions, so that six months later "wait, why did we choose that?" has an answer instead of a shrug.

It's not memory infrastructure. It's a second brain you can look at.

Scatterbrained

The new name was sitting there the whole time. Scatterbrained — for the people whose best thinking arrives in scattered pieces at inconvenient hours. The promise is right there in the joke: a mind full of scattered thoughts, wired together so they remember everything.

The shape of it:

Capture where it's natural. Jot in Notion, drop a file, paste a link — those are just on-ramps, and entirely optional. The graph is the canonical truth; the capture lanes are conveniences, not requirements.
Query where it's powerful. A real graph (Neo4j underneath), not a folder of notes. Follow a decision to the benchmark that drove it; follow a fact to the moment it was superseded.
Let an agent keep it honest. It ingests, links, and — crucially — invalidates instead of deletes, so the history survives the correction. (The one idea I'm keeping from the bi-temporal world, because it's simply the right way to remember.)

And the part I'm proudest of, the part that made the rename feel less like a loss and more like a promotion: an observatory. The whole graph rendered as a constellation you explore on your own machine. Click any node and the panel composes itself from what that node is — a goal shows its progress, a rule shows its confidence, a source opens its actual file, a decision shows the timeline of what replaced it and why. Nothing leaves your laptop.

Better for the fire

Here's the thing nobody tells you about being forced to rename: a good name is a constraint, and the right constraint makes the work truthful. "Engram" let me hide behind a category that wasn't really mine. "Scatterbrained" doesn't let me pretend it's infrastructure — it's a second brain you can see, owned and local and a little bit honest about being built by, and for, a scattered mind.

So: credit still where it's due — the original Engram is the real thing in the memory-engine lane; go read their paper. And the project that used to borrow that name is now Scatterbrained, finally wearing something that fits.

A phoenix, it turns out, doesn't rise despite the fire. It rises because of it.

— Ulric