Memory, custom instructions and personas

You should not have to restate your role, your stack and your preferred answer format in every new chat. This page sets up the three layers that carry across conversations — remembered facts, personalization, and a persona — and ends with a before-and-after test so you can see the difference in one sitting.

Cross-chat memory

Facts SynapseX remembers about you are injected into the system prompt on every turn, in every conversation. That is what makes a new chat start already knowing you work in TypeScript on a Postgres backend.

Where to manage it: open Context > Memories in the sidebar. The page lets you:

Add a memory by typing it into the input at the bottom
Search your memories
Delete any entry

The same list is reachable from Settings > Personalization through the memory management dialog.

The switch: Personalization has an Enable memory toggle. Turn it off and stored memories stop being used; turn it back on and they resume.

Asking the assistant to remember things

When the Memory tools capability is on — it is on by default — the assistant can remember, recall and forget facts itself. You do not need special syntax:


Remember that I always want SQL examples in PostgreSQL dialect, never MySQL.


Forget what you remembered about my previous employer.

Say it in plain language and check Context > Memories afterwards to confirm the entry landed or disappeared.

Personalization

Settings > Personalization holds the profile that is compiled into a block and prepended to your system prompt on every message.

Field	What it changes
Nickname	How the assistant addresses you.
Occupation	What you do — the assistant assumes your domain instead of asking.
More about you	Interests, values, constraints to keep in mind.
Custom instructions	Free text followed in every chat. This is the strongest lever on the page.
Base style and tone	Efficient, Balanced or Thorough — the default depth of an answer.
Warmth / Enthusiasm / Headers and lists / Emoji	Four tone traits, each set to Less, Default or More.
Fast answers	On: prefer direct answers when a request does not need deep work. Off: spend more effort on careful reasoning.

Custom instructions is where you put rules like “Always show the failing case before the fix” or “Never use bullet points in prose sections”.

Personas

The persona chip in the composer switches between five built-in system-prompt presets: a general default, a senior-engineering voice, a design-oriented voice, a research-oriented voice and a creative voice. Your choice persists in the browser.

A persona swaps the assistant’s role wholesale. Personalization tunes how it talks to you. They stack — persona, then personalization, then any enabled custom skills.

Inference parameters

The sliders popover in the composer sets three values applied to every message and stored in your browser:

Parameter	Default	When to change it
Temperature	0.7	Lower it (0.2–0.3) for code, extraction, maths and anything where you want the same answer twice. Raise it for brainstorming.
Top-p	1	Leave it alone unless you are deliberately experimenting.
Max tokens	4096	Raise it when replies are being cut off mid-answer.

Lowering temperature is the fastest fix for an assistant that keeps inventing variations of a factual answer.

The Capabilities switchboard

Settings > Capabilities has eight switches that genuinely change which tools the assistant can call. A new account starts with research tools, knowledge, memory, quantum, image, audio, agents, run_app and labs enabled.

In practice there are three you have a reason to touch:

Switch	Why you would change it
Memory tools	Turn off if you do not want the assistant writing or reading durable facts.
Research tools	Turn off if you want answers strictly from the model with no lookups.
Quantum QCOS	Keep on to ask about quantum backends — see Quantum backends.

Remember that a Turbo turn has no tools at all, whatever these switches say.

Worked example: see it change the answer

Open a new chat and ask:
```
What is the best way to handle background jobs in my app?
```
Note the answer: generic, hedged, probably asks what stack you use.
Fill in Settings > Personalization:
- Occupation: Backend engineer on a Node.js and Postgres API
- Custom instructions: Assume TypeScript. Always give a concrete recommendation first, then two alternatives with one line each on why not.
- Base style and tone: Efficient
Open another new chat and ask the exact same question.

The second answer names a specific approach for your stack, leads with the recommendation, and skips the “it depends on your requirements” preamble. Nothing about the question changed — only the block that gets prepended to it.

Custom skills — reusable instruction packs that stack on top of this.
Turbo — why memory tools are unavailable on a Turbo turn.