DMS - Dual Memory System - Character Tavern Docs

Overview

The Dual Memory System (DMS) powers conversation continuity by combining two complementary memory types:

Short Term Memory: The most recent messages the AI can see directly in its context window
Long Term Memory: Relevant memories automatically retrieved from your entire chat history

Long Term Memory is powered by compact memories - short factual statements extracted from your chat (for example: {{user}} likes pasta). These memories are what the system retrieves during generation.

DMS runs automatically in the background - no setup or manual management required.

Why it matters

Continuity: Keeps plots, relationships, and facts consistent over long sessions
Relevance: Surfaces the right past details at the right time
Zero setup: Works out of the box; you focus on the story

How DMS works (at a glance)

Ingest recent messages

DMS collects the latest part of the conversation to form Short Term Memory.

Select correct swipes

For each message with swipes, DMS picks the appropriate one for context.

Retrieve relevant history

When helpful, DMS searches your entire chat for semantically relevant memories.

Compose final context

Short Term Memory + retrieved Long Term memories are provided to the AI for generation.

Memory Details

Short Term
Long Term
Memories

Short Term Memory is the direct, visible portion of the conversation the AI can access.Trade-offs

Very large windows can dilute what is most important on smaller models
Messages outside the window are not directly visible
Larger windows can increase latency and cost on lower-end models

How it’s filled

DMS gathers the latest messages
Selects the correct swipes
Sends the last X messages within your token limit to the AI

Standard Short Term Limits by Tier

Tier	Tokens
Free	8,000 tokens
Premium Tier 1	10,000 tokens
Premium Tier 2	14,000 tokens
Premium Tier 3	18,000 tokens

Higher token limits allow longer recent context without relying on retrieval.

Long Term Memory preserves broader conversation knowledge via retrieval.Strengths

Recall details from any point in history
No context window limit
Core facts can persist and be reused

Caveats

Retrieval returns the most relevant info, not every detail
Extremely specific details may not match unless clearly related

Retrieval Process

Package recent context

DMS takes the last 5 messages from your conversation as the query context.

Send to RAG server

These 5 messages are sent to our retrieval (RAG) server to search stored memories.

Extract most relevant memories

The server selects the most relevant compact memories based on semantic similarity to the query context.

Inject into prompt

Selected memories are injected alongside Short Term Memory in the final prompt for generation.

Standard Retrieval Limits by Tier

Tier	Memories / 15 messages
Free	5 / 15 msgs
Premium Tier 1	10 / 15 msgs
Premium Tier 2	15 / 15 msgs
Premium Tier 3	20 / 15 msgs

Need bigger windows and stronger retrieval? Check out Memory Boost for boosted limits and pricing.

Memories are a core part of DMS. They are concise facts automatically extracted from chat messages and stored for retrieval.

Message saved or edited

When you save or edit a message, DMS schedules it for memory extraction.

Extract concise facts

The system converts relevant content into short, atomic statements.

Store in retrieval (RAG) server

Extracted memories are saved to our retrieval (RAG) server for fast semantic search.

Memories are intentionally brief, e.g., {{user}} likes pasta - compact, factual, and easy to retrieve.

Memory

​Overview

​Why it matters

​How DMS works (at a glance)

​Memory Details

​Standard Short Term Limits by Tier

​Retrieval Process

​Standard Retrieval Limits by Tier

Overview

Why it matters

How DMS works (at a glance)

Memory Details

Standard Short Term Limits by Tier

Retrieval Process

Standard Retrieval Limits by Tier