Skip to content

Long-Term Memory

Every model has a context window — a limit on how much conversation it can "see" at once. Long-Term Memory (LTM) keeps a running recap of the story so far and injects it into the prompt, so the model remembers chapter one by the time you reach chapter ten.

How it works

As a chat grows, Pyre periodically summarizes what's happened and stores it as a checkpoint. The accumulated recap rides along in the system prompt, compact enough to fit while preserving the threads that matter — who's who, what's been established, where the story stands.

The auto-summarizer fires every N messages (the threshold is configurable; set it to 0 to turn auto-summarize off globally). You can also summarize on demand at any time.

A story, not a synopsis

Each checkpoint is written as the next paragraph of one ongoing chapter, not a standalone bullet list. Concatenated, the recap reads as a single unbroken narrative — which keeps the model's continuation in tone, instead of feeling like it's reading meeting minutes.

Branch-aware

Pyre's chats branch: you can re-roll replies and rewrite your own turns without destroying the timeline that followed. Memory respects that. Each checkpoint is fingerprinted to the branch it was taken on:

  • Re-rolling a past message invalidates checkpoints for the new branch, but they stay valid on the original branch when you navigate back.
  • Stale checkpoints from abandoned branches sit harmless and hidden.

So your recap always reflects the timeline you're actually playing — not a tangle of every path you explored.

Staying in control

Action What it does
Summarise now Generate a checkpoint immediately, even with auto-summarize off.
Retry (per checkpoint) Regenerate a single summary you didn't like.
Edit Adjust a checkpoint's text by hand.
Delete Remove an individual checkpoint, or wipe all of them.
Per-chat toggle Turn memory on or off for one specific chat.

You can also tune the summary prompt template and the recap's length budget.

When something goes wrong, you'll know

If the summarizer hits a provider error (an auth failure, a timeout, an empty reply), Pyre surfaces it instead of silently giving up — so you can fix the root cause (often the model or key) rather than wondering why memory stopped.

Memory vs the other state tools

LTM is the narrative memory — the prose recap. If you also want structured state (locations, inventory, relationship beats tracked as fields) or authorial direction (planned story beats), those are separate, optional tools:

  • Live Sheet — a structured, auto-updating state panel.
  • Script — story-direction beats you want the model to steer toward.

See also