Skip to main content

Memory & Recall

This page explains the memory features that help the model remember across long conversations: the rolling conversation memory, keyword-activated memory entries, and vector recall.

Conversation Memory​

Every chat keeps a rolling summary that the pipeline maintains as the conversation grows. When the context-shifting strategy summarize is active, the oldest excluded history is condensed into a local extractive summary that is inserted before the current user input — so the model keeps the gist of early events even after the raw messages leave the token budget. The summary is stored with the chat and survives reloads.

You can see exactly what the current prompt contains before sending: a live context preview shows the selected tokenizer, the context limit and the reserved response space, excluded blocks, summarized blocks, and the strategy applied. See Settings for the strategy picker.

Memory Entries​

Memory entries are long-lived knowledge fragments that persist across chats, independent of any single conversation. Each entry has:

  • Scope — global or bound to a character.
  • Activation keywords — a case-insensitive substring match against the conversation context.
  • Content — the text injected when the entry fires.

This is the classic RAG pattern: retrieval is triggered by keyword match, and the injected fragments answer the model's need for stable facts — character details, world rules, or ongoing plot points — without bloating every prompt. Like lorebook entries, memory blocks are ranked by relevance in the prompt pipeline and count toward the token budget.

Vector Recall​

Vector recall is the context-shifting strategy vector-recall. Instead of cutting context purely by age, it ranks lorebook and memory blocks by semantic relevance to the current input and drops the least relevant ones first, then trims older history. The result: the model keeps the material that matters for the current message even when it is not the newest.

The strategy is selected per generation settings, and plugins can add further strategies through the SDK. Every strategy still respects the final host-controlled token budget — plugins cannot bypass it.

Choosing a Strategy​

The available strategies are truncate (drop oldest unprotected groups), summarize (condense excluded history), vector-recall (keep high-relevance blocks, trim by relevance and age), and manual (exclude specific messages from the prompt without deleting them from history). The manual mode exposes an action on each message to exclude or restore it, and tool-call/tool-result pairs are always handled together. See Chatting for message-level controls and Lorebooks for the related keyword activation model.