Your memory
day to day

Remember

Type a note in the composer and choose Save. Use #tags inline: the hashtags are removed from the text and stored as tags. A numbers-only hashtag such as #4821 stays in the text, because it reads as a ticket or pull-request reference rather than a topic.

Saved

Illustrative example

Launch the portfolio first. Add the shop later.

Recall

Ask a question, a topic or a fragment. Recall searches by meaning, builds an answer, and cites the memories it used. Open the source count to inspect the individual matches behind the answer.

Before you search, the Home screen can show recent capture volume and sources, a two-week activity strip, current topics, proposed patterns, and an older memory worth revisiting. A quiet brain stays quiet.

Recalled

Sourced from the saved example above

Launch the portfolio first. Add the shop later.

Browse

Memories has two views, and it remembers which one you used last:

• List shows memories newest first, with tag filters, and layer and author filters on a team brain. Each card leads with the memory itself, followed by source, time, tags and actions. Indexing indicators appear only while a memory is still being indexed, or when it cannot be found by semantic recall.
• Graph shows memories as nodes clustered by subject, with relationships as the lines between them. Select a node to open the full memory.

Open, edit, delete

Opening a memory shows the full content, the source, the capture time and last update, the tags, the importance score, the memory kind, the lifecycle status, the recall count, whether semantic recall has indexed it, and the memories connected to it.

• Append adds a timestamped update while preserving the original. Use it when something changes.
• Edit replaces the current content.
• Forget permanently removes the memory and its vectors. There is no undo.

Back up, restore, leave

Back up as JSON exports every memory and relationship you can read. Export as Markdown creates a readable archive. Restore is safe to repeat: imported memories always land in your personal layer, and restoring a backup never shares anything automatically. Afterwards, use Make searchable to rebuild any missing embeddings.

How recall behaves, in ordinary words

Every memory is stored as numbers that represent its meaning, not its wording. When you ask a question, your question gets the same treatment, and the brain finds the stored memories whose meaning sits closest to yours. So you can store "users drop off at the payment step" and later ask about "onboarding problems", and it surfaces. You never need to remember how you phrased something.

Freshness matters too. Results are re-sorted with a bias toward the new: a memory from today keeps its full score, one a week old keeps roughly ninety percent of it, two weeks old about eighty-five percent, and a month old around three-quarters. The fade slows down after that rather than continuing to zero. How often a memory has been recalled, its importance, and its record in contradictions also nudge the ranking. Old memories still surface when nothing recent is relevant; fresh versions simply win when the meanings are close.

How the brain reconciles, in ordinary words

Before anything is stored, it is compared by meaning against what is already there:

• Duplicate. A near-exact match is blocked outright and the existing memory is returned, while a probable duplicate is stored but flagged for you to review.
• Overlapping note. A clear overlap is merged into one current note, so nothing new is created.
• Changed fact. A contradiction is settled on the newer belief; the outdated memory drops out of recall, though its row is kept for audit.
• Protected memory. A canonical or high-importance memory is never quietly merged or replaced; it is flagged instead, so nothing critical changes without your awareness.

The technical ranking detail

A near-exact match is one at 95 percent similarity or more. A probable duplicate sits between 85 and 95 percent, stored but flagged as a duplicate candidate with a pointer to the memory it resembles. In a contradiction, the newer belief wins and is tagged as having resolved it, unless the older memory was marked canonical, in which case it wins instead and the newcomer is kept as a draft.

Two more quiet habits keep the brain tidy. Long notes are split into overlapping chunks before embedding, so every part of a long note stays searchable rather than the tail going invisible. And a nightly job compresses the small, repetitive notes that gather under a busy tag into one digest, while anything important, anything you have recalled repeatedly, or anything that has survived a contradiction is kept whole.