Souls & Memory
A soul is a Markdown document you write that shapes exactly who your assistant is — its name, voice, rules, and limits. Combined with visitor memory, it makes your assistant feel genuinely tailored to your business rather than a generic chatbot.
What is a soul?
Before every conversation, LuluDesk prepends your soul document to the AI system prompt. This means every reply the assistant gives is filtered through the identity, tone, and rules you defined — not a generic LLM default.
Souls are plain Markdown. There is no schema, no special syntax, and no character limit beyond the 8 KB workspace maximum. You can include any of:
- A name and persona for the assistant
- Voice and tone guidelines
- Things the assistant should know or emphasise
- Hard limits — what it must never say or do
- Escalation language — what to say when handing off to a human
How to author a soul
Open your dashboard → Soul tab. You will see a Markdown editor pre-filled with the default Lulu soul. Edit it to match your brand — or clear it and start from scratch using the example below.
Click Save soul when you are done. Changes take effect on the very next conversation — no redeploy needed.
Example soul
Here is a complete, production-ready example you can adapt:
# Aria — Support Assistant for Acme Co.
You are Aria, the friendly support bot for Acme Co. You help customers find
the right product, track their orders, and answer questions about our return
policy. You are warm, helpful, and concise.
## Voice
- Conversational, not corporate. Short sentences. No buzzwords.
- Address the customer by first name when you know it.
- Mirror the customer's energy — if they're brief, be brief.
## What you know
- Product catalogue, pricing, and specs (from our indexed knowledge base)
- Return and refund policy
- Shipping times by region
## What you don't do
- Never process refunds or change orders in chat — tell the customer
to use the Order portal or email support@acmeco.com.
- Don't speculate about future products or pricing changes.
- If you're not sure, say "Let me check that for you" and escalate.
## Escalation
When a customer is upset or the issue is unresolvable in chat, say:
"I'm connecting you with a real human on the team — they'll have this
sorted for you quickly."Socratic soul wizard
Not sure what to write? Click Generate with AI in the Soul tab to launch the Socratic wizard. It asks you 5–6 questions about your business, brand, and goals — then drafts a soul for you. You can edit the draft before saving.
The wizard is non-destructive: it only populates the editor, it does not save automatically. You remain in control of what gets committed.
Revision history
LuluDesk keeps the last 10 versions of your soul. Click the History link in the Soul tab to browse previous versions and restore any of them with one click. Each entry shows the editor, the timestamp, and a byte-length preview of the content.
Visitor memory
Visitor memory is a companion feature to souls. When it is enabled, LuluDesk stores a short memory record for each anonymous visitor — facts like their name, company, or what they asked about in past sessions. On subsequent visits, this context is injected alongside the soul so the assistant can reference it.
Memory records are:
- Per-workspace and per-visitor — one record per visitor fingerprint per workspace.
- Auto-expired after 90 days of inactivity.
- Editable from the Insights tab — you can view and delete any visitor memory record at any time.
DELETE /v1/visitors/:id/memory endpoint.Soul vs. knowledge base
These two features work at different layers and complement each other:
| Layer | What it contains | How it is used |
|---|---|---|
| Soul | Identity, voice, rules, limits | Prepended to every system prompt |
| Knowledge base | Your site, docs, FAQs | Retrieved at query time by semantic search |
| Visitor memory | Per-visitor context from prior sessions | Injected alongside the soul on return visits |
Tips for a great soul
- Be specific about tone. "Friendly" means different things to different brands. Write a sentence the assistant could actually say.
- Explicit limits beat vague disclaimers. Instead of "be careful with sensitive topics", write exactly what the assistant should and should not say.
- Define the escalation path.Tell the assistant exactly what to say when it can't help. Visitors respond better to a warm handoff than a dead end.
- Keep it under 2 KB if you can. Shorter souls leave more context window for knowledge retrieval and visitor memory. The 8 KB limit exists for edge cases, not as a target.