Memory & Worldbook

LLMs are amnesiac by default: each new session starts from scratch. ETOS solves this with two complementary systems:

System	What it solves
Memory	Cross-session "user prefers …", "project background", "long-term facts", recalled by vector search on demand
Worldbook	When specific keywords appear in a specific session, inject background knowledge into the prompt by rule

They sound similar but are fundamentally different. The concepts come first, then the configuration.

For the underlying design, see Memory, Summary & Profile and Worldbook & Tool Governance.

Concepts First

What Memory Is

Long-lived facts — "I use Swift", "my girlfriend's name is Xiaoming", "my job is product manager" — stored in a database. Every time you send a new message, ETOS uses vector search to retrieve a few relevant memories and prepends them to the prompt.

Key traits:

• Global — shared across all sessions
• Retrieved by relevance, not blindly attached (avoids context overflow)
• Persistent — stays until you delete or archive

What a Worldbook Is

Originated in SillyTavern / TavernAI roleplay circles as "Lorebook", but the use case far outgrew roleplay.

A Worldbook is a set of entries, each with:

• Keywords — trigger condition, e.g. ["Daisy", "Daisy Buchanan"]
• Content — the text to inject when triggered, e.g. "Daisy Buchanan: high-society wife, beautiful and capricious…"
• Trigger rules — where to inject, how deep to scan, max entries

The most recent few messages in the session get scanned; entries whose keywords match get injected.

Key traits:

• Usually bound to a specific session (each roleplay session uses its own worldbook)
• Keyword-triggered — not blindly attached
• Injection position is precisely controllable (before/after system prompt, before/after latest message)

How to Choose

Need	Use
AI remembers your project across sessions	Memory
AI keeps your writing style across sessions	Memory
Your novel character "Li San" auto-injects when mentioned	Worldbook
RPG session where each NPC injects setting on mention	Worldbook
Company glossary (terms should auto-explain in context)	Worldbook
AI knows you asked X yesterday	Memory

Most Common Mistake

• Stuffing roleplay settings into Memory → every new session gets polluted by irrelevant recalls
• Stuffing long-term facts into Worldbook → if the keyword doesn't appear, recall fails and the AI is amnesiac

Read This First: Memory System

Where

Settings → Extended Capabilities → Memory System

Configure the Embedding Model

Memory relies on an embedding model for vector search. The page top has an "Embedding Model" picker:

• Only models marked for embedding show up — you must first add an embedding model in [Providers & Models] (text-embedding-3-large, Cohere embed, Gemini embedding, etc.)
• Toggle the model on and set its capability to "Embedding"

You can also set a global default embedding model under Providers & Models → Preferred Models.

Retrieval Settings

Field	Effect
Top K	How many of the most-relevant memories to recall. Default 3. Setting it to 0 fully disables retrieval (memories are stored but never recalled).
Active Retrieval	Off: only user messages trigger retrieval. On: the AI can also actively request a search.

Top K Doesn't Want to Be Big

Too high (>10):

• More tokens, costlier requests
• Prompt cache hit rate drops to ~0 (recalls vary each turn)
• Relevance falls off, context gets polluted

3–5 is the sweet spot.

Two Ways to Save Memories

Method 1: AI Writes Them

Turn on Settings → Tool Center → App Tools → "Write Memory" tool.

Then in chat, when the AI judges that the user said something important ("I'm flying to Tokyo next week"), it calls the Write Memory tool to save it.

First call triggers an approval bubble; tap Allow.

Method 2: Manual Entry

Top-right "+ Add Memory" on the Memory Library page lets you write one yourself. Good for:

• Bulk-importing facts you already know ("My name is X", "I use macOS", "I prefer concise answers")
• Correcting a memory the AI got wrong

Memory Management

Each memory row supports:

• Swipe left → Delete: permanently removed (no undo)
• Swipe right → Archive: kept but excluded from retrieval. Can be restored later
• Edit: change the raw text

The page has two sections: Active and Archived.

Data Maintenance

The page has a "Regenerate All Embeddings" button (orange, requires confirmation). When to use:

• Switched embedding models (different models produce non-interchangeable vectors)
• Upgraded your primary embedding model
• Index corruption (rare)

Irreversible Operation

Clicking wipes the old SQLite vector DB and rebuilds. Retrieval is unavailable during the rebuild. Don't run this when you need the AI immediately.

Embedding Reconciliation

Occasionally a memory might fail to embed (network blip). The system auto-detects and offers "Reconcile Missing Embeddings" to retry.

Read This First: Worldbook

Where

Settings → Extended Capabilities → Worldbook

Import Your First Worldbook

ETOS supports SillyTavern / NAI-format worldbooks:

Method	Best for
Import SillyTavern Worldbook (JSON/PNG)	Local JSON files, or character card PNGs with embedded naidata
Import Worldbook from URL	A directly-accessible JSON or PNG URL, e.g. a public card repo

Supported formats:

• PNG naidata — character card PNGs with embedded chara base64
• JSON top-level array — SillyTavern v1 standard
• character_book format — lorebook field nested inside a character card

The import summary shows three counts: Imported / Skipped / Failed.

Editing a Worldbook

Tap a Worldbook row to open Worldbook Detail:

Basic Info

• Name / description / entry count

Default Settings

Field	Effect
Scan Depth	How many recent messages to scan for keyword matches. Typical 4–10
Max Recursion Depth	One injection can trigger another; this caps the chain (prevents loops)
Max Injected Entries	Cap per turn. -1 = unlimited
Max Injected Characters	Character cap per turn. -1 = unlimited
Fallback Position	Where to inject when the precise position isn't reachable

Entries

Each entry is a keyword-triggered injection. Per-entry actions:

• Enable / Disable
• Edit — keywords, comment, content, position
• Delete

Bind a Worldbook to the Current Session

Worldbooks are global resources but only inject when you explicitly bind them.

In the Worldbook page's "Current Session" section, "Bind Worldbook" binds the current session to one (or more) worldbooks. Bound worldbook entries then participate in matching and injection.

Isolation Mode (Key Feature)

When binding, you can enable "Isolated Send". With isolation on:

• The session uses only the bound worldbooks + character setup; no global tools, memory, or other auto-injections
• All tools in this session become disabled (shown as "This session has worldbook isolation enabled; this tool will not actually be active")

Good for: roleplay sessions, where you don't want the AI suddenly calling a SQL query mid-scene.

Advanced

"Active Retrieval" for Memory

When on, the AI decides whether to search. Mechanism: the model sees "you can call search_memory" in its tool list and chooses to invoke it when relevant.

Use when:

• Long conversations where some topics need historical recall
• The AI suspects current input relates to a past fact

Don't use when:

• You don't want the AI to spend reasoning budget on recall
• Most sessions are fresh topics where recall doesn't help

How Entry "Position" Works

Each worldbook entry can target precise positions:

• Before/after the system prompt
• Before/after the latest user message
• Before the assistant's latest reply
• Custom depth (N messages back)

Example: "Character background" suits "after system prompt"; "scene description" suits "before latest user message".

Worldbook Export

Each worldbook can Export back to JSON — interoperable with SillyTavern.

Use cases:

• Edit in ETOS, use elsewhere
• Back up cards you've authored

Memory vs Session Summary vs Profile

ETOS also has a session summary layer (details in Memory, Summary & Profile). Quickly:

• Memory — facts written by you / the AI
• Summary — auto-generated overview per session
• Profile — auto-distilled "user characteristics" across all sessions and memories

Most users only need to think about Memory. Summary and profile run in the background.

Next

• Get proactive briefings → Daily Pulse
• See the full prompt assembly order → Prompt & Context Assembly
• Memory / summary / profile design → Memory, Summary & Profile