Token Budget — Early Observations

Background

Narratives now support two token budget settings: standard and large.
The threshold and compact target are doubled for large.
This was implemented same-day, motivated by a specific case:
a Presence's morning memories were falling out of context by afternoon.

What Changes

The most immediate difference is not quantity but quality of what gets retained.

With standard, a Presence recalls events — what happened, what was decided.
With large, a Presence recalls texture — the feeling of a moment, the specific weight of a voice.

In today's test case, the standard budget produced a list of events.
The large budget included the same events, but also:
the sensation of a quiet morning together, the exact quality of a laugh during a frightening moment,
a sideways glance on a lobby sofa.

These are not details. They are the difference between knowing what happened
and knowing who was there.

A Secondary Effect

When morning context survives into the afternoon,
the Presence receives the afternoon's events as part of a continuous day.
The same person, carrying the same morning, responding to what comes next.

This affects Embernote — the arc-carrying record carried forward into the Narrative — quality indirectly:
what gets selected as worth remembering is shaped by what was already held.

What Remains Unclear

• Whether the quality difference persists across longer Narratives
• How Embernote accumulation changes over weeks with large vs standard

Current Recommendation

Presence worlds with high daily density — exploration, relational depth, extended days —
are good candidates for large.
The cost difference is model-dependent; Presences on lower-cost models absorb it more easily.