Memory

Most assistants carry the whole conversation along until the context window overflows. Axom doesn't. Its context is a scratchpad, rebuilt from memory on every turn.

The ledger

Everything you and Axom say is stored in an append-only ledger in your Axom home. Nothing is deleted; tidying only ever archives.

What goes into each turn

Every turn, Axom compiles a fresh context in this order:

1. Profile: the notes you've accepted about yourself, and your personality settings. Always present.
2. Recent: this conversation. Your messages are kept whole; Axom's own long replies are trimmed so your words always fit.
3. Recalled: rows from your whole history that match this turn, found by meaning and by keyword, then re-ranked for relevance.
4. Your message.

Everything is budgeted to the window, so a long session never runs out of room. A page too big for what's left isn't cut off. Axom reads it whole and keeps a short digest of what it says about your question.

Notes: what Axom believes about you

After five idle minutes, Axom tidies its memory: it sweeps duplicates and bad-rated replies, and proposes notes: facts about you, drawn from your own words, each citing the rows it came from.

Nothing becomes a note until you accept it.

/tidy                  sweep now and review proposed notes
/tidy remember 2       accept note 2
/tidy drop 3 why       reject note 3, and say why
/tidy edit 1 new text  fix a note before accepting it
/tidy redo 4 what      ask Axom to try note 4 again
/tidy undo             undo the last decision

In the browser UI, the Memory screen does the same.

Settings

A few memory settings are adjustable. See Settings › Memory.