Engram

Engram is a memory I share with my agents. I hand it anything worth keeping, and it keeps it. It files what it can on its own, and it asks before it changes anything it is unsure about. Later I ask a question, and it answers from what I gave it, with sources.

Why it exists

Chat assistants are good at reading and bad at remembering. I once gave a chat a stack of reports. It summarised them well, inside the conversation. The memory behind it got a pile of flat entries, with no type and no links. The understanding left with the session.

A folder of notes would not fix that either. Something has to do the filing, check its work against what is really stored, and ask me when it is unsure. That is what Engram is for.

Keeping

The moment of keeping should cost nothing. I paste text, share a link from my phone, or drop in a recording. Engram answers at once. The reading happens after, in the background: it pulls the article out of a page, the words out of a PDF, and a transcript out of a voice memo.

Filing, and asking

An agent reads each new item and suggests what it is: a type and tags, links to related items, and duplicates to merge. It cannot write those changes directly. Each suggestion is a proposal. Most proposals wait for me in one list, and a few confident ones apply on their own.

I go down the list and accept, edit, or reject. Every applied change can be undone. Each decision is recorded, so the organizer can learn from it later.

Finding

I search by meaning or by the words I remember, and a near-miss in spelling still lands. Answers come with the items they used, so I can check them.

Two promises hold underneath. An entry I tag no-model is never sent to any model, but it is still kept and still found. And Engram can write my entries and links to a git repository I own, so I can take them with me.

Where it goes next

The first version is live. In my own words from the design: "my belief was always that v1 Engram is all about storage. v2 is about how intelligent that storage can become."

The organizer already reads, files, and asks, after each capture and again every night. Each night it also checks that each file an entry mentions is really stored. Next it will compact old material into summaries, and it will learn my way of filing from each decision I make. Small apps sit on top, one part of life each: Health is being built, and Finance is in design. The system I use to build software already runs on Engram.