How memory works

Most tools that "remember" your chats keep the last few messages. Selah keeps something smaller and more durable: individual things you appear to believe, each with its own history, confidence, and links back to where you said it.

That's why the companion can tell you what you thought about a passage two years ago — and notice that you've since changed your mind.

Facts, not transcripts

Selah doesn't store your conversations as a wall of text and search it. It pulls out atomic facts — one claim each, small enough to be right or wrong on its own.

A fact records what you said, how confident Selah is that you meant it, how many times you've said something like it, and which notes it came from. Facts attributed to someone else are marked as theirs, so a teaching you recorded from a conference speaker never gets read back to you as your own conviction.

A belief has versions

This is the part most memory systems skip. When you write something that contradicts what you said before, Selah doesn't overwrite the old view, and it doesn't keep both as equals. It supersedes.

Three things follow from that:

• Saying something twice makes it stronger. Repetition raises confidence rather than creating a duplicate.
• Changing your mind is a first-class event, not a conflict to resolve. The new view becomes what Selah works from immediately.
• Nothing quietly rots. A superseded view leaves recall the moment it's replaced, so it can't resurface later and contradict your current position. It stays linked to the version that replaced it, so the trail of how your thinking moved survives.

Some facts are ephemeral by design — "preaching on Romans this Sunday" expires on its own rather than hanging around as though it were a conviction.

How Selah decides what to bring back

When you ask the companion something, it does not hand over your most recent notes. It scores your facts against what you actually asked, on six signals at once:

Bar length is proportional to real influence.

Recency sitting near the bottom is the whole design. For a study habit measured in years, the useful thing is almost never the latest thing — it's the thing that bears on the passage in front of you. A note from four years ago about Romans 8 should beat yesterday's unrelated thought about church logistics, and it does.

The highest-scoring facts then compete for a limited amount of room in the conversation, so the companion gets a short, strong set rather than everything that matched.

Facts that are superseded, expired, or flagged as suspicious are excluded before scoring starts. That's a hard filter, not a low score: no combination of signals lets a replaced belief back in.

If Selah can't work out what you're asking, recall falls back to your most recent active facts rather than returning nothing.

Everything traces back

Every fact records which notes it came from. Delete the note behind a belief and the belief goes with it. Nothing is remembered from nowhere, so you can always check what you actually wrote and disagree with your past self.

Positions that reach your knowledge map follow the same lifecycle — supersede the fact and the claim on the concept moves with it.

The switch, per note

Every note has a Use in Selah's memory toggle.

Turn it off and that note stops feeding recall entirely. It stays saved, it stays searchable, and you can still read and edit it. Anything Selah had already derived from it is removed as well — not hidden, removed.

Use it for a draft you're writing for someone else, a private journal entry, or anything you'd rather keep out of the conversation.

What memory never does

• It doesn't train an AI model. Your study content is never used for training, by us or by our model provider — see the privacy policy.
• It doesn't cross accounts. Memory is scoped to you. No part of one person's study can reach another's.
• It doesn't take orders from your own notes. Saved content is treated as material to think about, never as instructions to the companion.