Three bounds

Messages, minutes and characters, tightest binds

Three independent limits on how much of a thread reaches her prompt, with the ranges and defaults the console ships.

SettingWhat it boundsRangeDefault
Messages she can see Most-recent group messages (not turns) in her prompt; 0 = memory off 0-100 20
How far back (minutes) Age window; the tighter of count and window wins 0-720 30
Hard character budget Ceiling on assembled history in characters; oldest dropped first 0-8000 4000

Labels, ranges and defaults quoted from site material Block 3 section 3 · the Interaction memory page, its labels, ranges and shipped defaults

Why three and not one

Three independent limits, and the tightest one binds. Messages, minutes and characters bound different things on purpose: a busy minute and a quiet hour are not the same amount of conversation, and neither is a hundred one-word messages and twenty long ones. Setting the message count to zero turns memory off entirely.

All three are deployment-wide rather than per bot, so two bots in the same community cannot disagree about how much of a thread they can see.

What goes first

When the character budget binds, the oldest messages are dropped first, so what survives is the part of the conversation nearest to what is being answered.

The thread

What she can see, and what she never can

The group's conversation, her own replies marked as hers, and four things that never enter the prompt at all.

What reaches her

The whole group thread, including her own replies, which are marked as hers so she does not read them back as somebody else's words. It is the group's conversation she sees, not a private per-member history.

What never does

  • Destroyed messages.
  • Deleted messages.
  • Messages a moderation decision rejected.
  • Anything at all from a member who revoked their consent.

That last one is the load-bearing case, and it is the same rule the rest of the system runs on: revocation is evaluated on every read rather than applied as a one-off cleanup, so a member who withdraws does not merely stop appearing in the archive, they stop being in what she can see at all.

Publication is a view evaluated on every read, so a revocation removes a member from every public surface and from the assembled history at the same moment. built

Check /en/plugins/archive

The cost

What memory costs the prompt, measured not estimated

The console assembles a real prompt and prints what it costs against the window, warning before the window runs out.

Memory is not free, and the console refuses to let an operator guess what it costs. The cost card assembles a real prompt for the selected bot through the reply path's own function and prints three figures: the characters her rules, voice and facts take on their own, the characters with a full history at the current settings, and roughly what that is in tokens against the served window of 24,576 tokens SERVED_CONTEXT_TOKENS, src/interaction/reasoning.ts:147 (D-231, measured) · host OLLAMA_CONTEXT_LENGTH. Above 73% of the window it warns.

It measures rather than estimates, which is why it is the number to trust over anything printed on this site: the figures here were gathered once, and that card is assembled fresh for the bot in front of you.

A longer memory is not a better bot. Every character of history is a character the model reads before it reaches the question it was asked, and the window it all has to fit inside is fixed.

What that window is, and what a prompt costs before any history is attached, is on the system prompt page.