Vectorization & Semantic Memory
Deep dives into every tool on stage
Vectorization & Semantic Memory
Vectorization is the quiet retrieval layer behind parts of RoleCall that need to find the relevant piece of a large body of text instead of sending the whole thing to the model every time. It turns text into embeddings: compact numeric representations that let the app search by meaning, not only exact words.
For example, a query about “the debt the innkeeper is hiding” can retrieve an entry that says “Mara owes the guild three months of protection money,” even if the query never uses the same words.
What It Powers
RoleCall uses semantic retrieval for material such as:
- Lorebooks and Compendium entries — so a scene can retrieve the most relevant people, places, facts, items, and events instead of loading an entire world bible.
- Character-memory excerpts — so durable, important facts from a scene can be found later by meaning.
- Other opt-in memory and knowledge features that explicitly say they are using semantic search.
Vectorization does not increase a model's context window, change what a model is capable of writing, or replace your original content. Your source text remains the source of truth; embeddings are a search index that helps select useful context.
What “20 MB” or “40 MB per Month” Means
The allowance is a monthly vectorization-usage budget: an estimate of how much source text RoleCall can submit for embedding during that billing period. It is not a count of memories, a chat-history limit, a file-upload limit, or a gauge of permanent storage remaining.
As a deliberately rough planning estimate, RoleCall's quota calculation treats 1 MB as about 250,000 tokens of source text. That means:
| Monthly allowance | Very rough source-text equivalent |
|---|---|
| 20 MB | about 5 million tokens |
| 40 MB | about 10 million tokens |
Treat those figures as orientation, not a promise. Tokenizers, language, formatting, duplicate detection, and the exact content being indexed all affect real usage. Most people will never think in megabytes: the meaningful distinction is that a larger allowance lets you index more or larger lorebooks and preserve more memory excerpts in that month.
The allowance resets with the billing cycle. Text already indexed remains usable for retrieval; the quota controls new indexing work. If you exhaust it, RoleCall tells you rather than silently pretending the new text was indexed. For example, a manual memory excerpt is declined until the next reset instead of being stored as if semantic search could find it.
What Happens When You Add or Change Content
When an eligible feature indexes text, it derives an embedding from the text and stores that search representation alongside the source. Later, when the scene needs relevant context, it compares the current query with those embeddings and selects the closest matches.
That has a few practical consequences:
- Longer source material uses more of the monthly allowance. Keep lorebook entries focused and structured; one well-scoped entry is easier to retrieve than a giant catch-all page.
- Changing content can require fresh indexing. Treat substantial edits as new retrieval work, because the old embedding describes the old text.
- Duplicate or empty material should not become useful memory merely because it was submitted. The system can skip work that has nothing meaningful to add.
- Retrieval is selective, not clairvoyant. Good names, clear facts, and distinct entries help the system fetch the right context.
BYOK and Billing
BYOK means a supported call can be billed by the provider account whose key you connected, rather than by RoleCall's hosted allowance. It does not turn an external provider's usage into “free” — the provider bills you under its own terms.
The plan and model picker show the current availability and cost path for your account. If you are deciding between an included allocation and BYOK, use that live product information rather than treating the MB figure as a permanent storage entitlement or a fixed provider price.
Practical Advice
- Start with the material that must survive long scenes: setting facts, recurring NPCs, obligations, inventory, and important relationship changes.
- Keep one fact or tightly connected cluster per entry. Specific entries retrieve more reliably than an encyclopedic wall of text.
- Do not vectorize ordinary short chat turns just to be safe. Use excerpts or lore entries for information you actually want found later.
- If retrieval is not working as expected, first check whether the text was indexed and whether the current query gives the system enough distinctive detail to match it.
See Compendium for organizing world knowledge, Story Director for scene-level memory tools, and Image Generation for the separate image-credit system.