> ## Documentation Index
> Fetch the complete documentation index at: https://docs.dialtu.com/llms.txt
> Use this file to discover all available pages before exploring further.
# Contact memory
> What the agent remembers and what it contributes to shared memory
Dialtu keeps a **per-contact memory**, shared across channels: what a WhatsApp agent learns is
available to your calls, your SMS and your other agents.
The **AI Memory** group controls two separate things: what this agent **reads**, and what it
**writes**.
**Reading is opt-in; writing is on by default.**
*Use memory* is **off** by default, but *Record to memory* is **on** — on every agent, new and
existing. If you've never opened this section, your agent is already contributing everything
it learns to the contact's shared memory.
**📷 Screenshot needed** — the AI Memory group expanded, with both toggles and the Configure link.
Save to `/images/en/whatsapp-agents/memory-group.png`
## Reading: what the agent knows
### Use memory
Off by default. Turning it on gives the agent what's already known about the contact: facts,
preferences and open items gathered from WhatsApp, calls, SMS and notes.
### How the agent gets the memory
| Mode | What it does |
| ----------------------- | --------------------------------------------------------------------------------------------- |
| **Automatic** (default) | Dialtu injects the memory into the prompt, with rules stopping the agent **reciting it back** |
| **Variables** | You decide where it goes, by placing `{{memory_...}}` tokens in the prompt |
In **Variables** mode, if you place no tokens, **the agent gets no memory at all** — and
there's no visible warning in the editor. If you pick this mode, place at least one.
The nine available tokens:
| Token | What it inserts |
| --------------------------- | ---------------------------------------- |
| `{{memory}}` | Everything remembered, as one block |
| `{{memory_one_line}}` | A single sentence summing the contact up |
| `{{memory_summary}}` | The rolling summary |
| `{{memory_facts}}` | Concrete facts, as a list |
| `{{memory_preferences}}` | How they like to be treated |
| `{{memory_open_items}}` | What was left pending |
| `{{memory_journey}}` | Recent milestones |
| `{{memory_last_seen}}` | How long since the last contact |
| `{{memory_last_sentiment}}` | How the last conversation felt |
**For a brand-new contact, the tokens reach the model written literally.** If the contact has
no memory yet — every first-time customer — a prompt saying `Context: {{memory_facts}}` sends
exactly that text to the model.
This doesn't happen in **Automatic** mode. It's a strong reason to prefer it unless you need
fine control.
### Deep recall
Off by default, nested under *Use memory*. Lets the agent **look up earlier conversations**
mid-chat when the customer mentions something specific.
* Searches WhatsApp, voice and SMS history, up to **18 months** back.
* **Costs no extra credits**, but replies take a little longer.
Turning off *Use memory* also turns off deep recall, and re-enabling it **doesn't restore**
it.
## Writing: what the agent contributes
The **Record to memory** toggle (on by default) and its **Configure** link open the *What this
agent records* modal.
**📷 Screenshot needed** — the "What this agent records" modal, showing the four categories.
Save to `/images/en/whatsapp-agents/memory-modal.png`
### What to record
Four categories, **all on** by default:
| Category | What it stores |
| --------------- | ------------------------------------------------------------------------- |
| **Facts** | Concrete, durable details: what they bought, who signs off, their address |
| **Preferences** | How they want to be treated: language, channel, schedule, tone |
| **Open items** | What was left unresolved, so the next agent can pick it up |
| **Milestones** | Notable moments: a quote sent, a visit booked, a purchase |
Each one takes its own instruction (max 300 characters) explaining **what counts as that, for
this agent**.
### Context and exclusions
* **Business context** (max 600 characters) — shared context for every category. For example:
*"We sell wholesale to shops, not to the public."*
* **Never record** (max 600) — a **hard exclusion**: nothing matching it is stored, even if the
customer states it outright. This is where card numbers, national IDs and health data go.
Fill in **Never record** before enabling memory in a regulated industry. It's the only hard
barrier.
### When to record
| Option | When it writes |
| ------------------------------------- | ---------------------------------------------------------------- |
| **During the conversation** (default) | After roughly 30 minutes **of silence**, and again on resolution |
| **When it ends** | Only on resolution or expiry |
*During the conversation* means 30 minutes of **inactivity**, not every 30 minutes. A
continuously busy chat writes nothing until it goes quiet or is closed.
## Things worth knowing
Memory is read at the start. For the session's 24 hours, whatever the customer says
**doesn't appear** in that agent's own memory. It looks like a bug in the test panel, but
it's deliberate.
Conversation sentiment lives on the same record. Switching off *Record to memory* stops it
being captured — and the interface doesn't mention it.
These settings apply to **this WhatsApp agent** only. Extraction from voice calls and client
notes always happens, under its own rules.
Memory has a size budget and a maximum age. "The agent forgot something I can see on the
contact's memory page" is expected behaviour, not a fault.
It can't be switched off for one agent: it's rebuilt from what every agent contributes. What
you change here is **what this agent feeds into it**.
## Next step
What happens when the customer stops replying.