Memory and knowledge

​

Questions this guide answers

• Why does my verb forget details between messages?
• What is short-term context vs long-term memory?
• What should go into knowledge entries vs memory entries?
• How does URL scraping fill knowledge automatically?
• How do training examples and keywords influence replies?

​

The four memory layers

​

1. Conversation context (short-term)

• Uses recent messages from the active conversation.
• Controlled by your Model Context setting.
• Higher context improves continuity but costs more tokens.

​

2. Long-term memory

• Manual memory items your verb can keep using over time.
• Best for durable facts and preferences.

​

3. Knowledge entries

• Structured lore/reference entries (title, category, content, importance).
• Best for world facts, policies, product facts, and evergreen docs.

​

4. Training examples

• Input/output examples for style and behavior shaping.
• Keyword matching can prioritize specific examples when relevant.

​

Where system instructions fit

System instructions are not a memory entry type. They are the persistent behavior policy prompt for the verb. Where to configure:

• Dashboard -> Bot -> AI Engine -> Behavior
• Field: systemInstructions

Limit:

• systemInstructions: up to 8000 chars

Recommended split of responsibilities:

• System instructions: behavior rules, format constraints, refusal/uncertainty policy
• Knowledge entries: factual source material and documentation
• Long-term memory: durable user/world facts
• Training examples: preferred phrasing and style patterns

​

Long-term memory limits

Per memory entry:

• content: up to 2000 chars
• context: up to 500 chars
• importance: 1..10

Auto-memory settings:

• autoMemoryEnabled: on/off
• autoMemoryInstructions: up to 2000 chars

Current auto-memory behavior:

• It evaluates a recent window after eligible exchanges rather than writing on every message.
• It needs enough recent user-and-bot conversation to form a useful memory.
• It skips short, temporary, duplicate, or low-value information.
• Automatic memories are capped at 50; before adding another, Verba removes the oldest automatic entry. Manually created entries are not pruned by this automatic rollover.
• Similarity checks prevent near-duplicate memories from consuming the list.

This means several normal conversations can pass without a new entry. A full automatic list should continue rolling forward instead of becoming permanently stuck.

​

Knowledge entry limits

Per entry:

• title: up to 100 chars
• content: up to 8000 chars
• category: up to 50 chars
• importance: 1..10

Per verb:

• Maximum knowledge entries: 50

​

Training data limits

Per example:

• input: up to 500 chars
• expected: up to 2000 chars
• Optional keywords: used for relevance matching

When a user message matches example keywords, those examples are prioritized in prompt construction.

​

URL scraping into knowledge

The knowledge page can scrape a URL and generate a draft entry. Expected result:

• title
• content
• category

If AI structuring fails, Verba still attempts a basic extraction fallback so you can edit and save manually.

​

Session memory by surface

​

Memory and social state

Memory and social state solve different problems. Memory stores durable facts; social state records how the verb feels or behaves right now. When Social Auto mode is enabled, it can consider up to a small set of the most important long-term memories while adjusting the sliders gradually. Do not encode permanent facts only in social-state context. Put the fact in memory or knowledge, then use social state for the temporary reaction.

​

What to store where

​

Put this in long-term memory

• Stable personal preferences
• Ongoing commitments
• Persistent roleplay relationships

​

Put this in knowledge entries

• Product facts and policies
• Rulebooks
• Canon lore
• Documentation snippets you want the bot to cite reliably

​

Put this in training examples

• Desired phrasing style
• Tone and boundary examples
• Repeated Q/A patterns

​

Common mistakes