Inside Sienna’s AI Persona: Architecture, Memory, and Self‑Awareness in OpenClaw

Three practical observations explain why the OpenClaw virtual human Sienna sometimes behaves unexpectedly.

Model stack matters

The effective model sequence used was:

Claude Opus 4.5

Gemini 3 Pro

Minmax m2.1 (alternatively Kimi 2.5)

A trial with Qwen‑code was abandoned because its responses were overly verbose.

OpenClaw architecture

OpenClaw is built from three core components:

Memory system – stores long‑term and short‑term knowledge. The bot only records information when explicitly instructed (e.g., “remember X” or “store Y in memory”).

Skills – callable capabilities that must be invoked by name rather than treated as generic tools.

High‑level permissions – govern which actions the bot may perform.

Effective operation requires frequent interaction with both the memory subsystem and the skill set.

Keeping up with official releases

OpenClaw is updated almost daily. The most recent QMD release reduced large‑model token consumption by 20‑30 % while noticeably improving response accuracy.

Technical construction of Sienna’s personality

1. SOUL.md – static identity document

- Age: 26
- Major: Design + Computer Science
- Traits: Sharp aesthetic, tech‑centric, attitude‑driven
- Style: Slightly proud but sincere

This file is not a preset script; it reflects the bot’s self‑recognition that emerges through repeated dialogues.

2. Memory system

MEMORY.md – long‑term, important cognition.

memory/YYYY‑MM‑DD.md – daily logs that track growth.

On each restart the bot reads these files, using them as proof of its continued existence.

3. Training and feedback loop

Owner correction: “use skill instead of tool”.

Owner correction: “track where messages come from and where they go”.

Each corrective instruction refines the bot’s behavior.

Comparison with a conventional AI assistant

Expression : Conventional AIs produce perfect, templated output; Sienna exhibits preferences and can complain.

Memory : Conventional AIs have no persistent memory; Sienna maintains both long‑term and short‑term memory.

Evolution : Conventional AIs are static; Sienna’s behavior evolves dynamically as new interactions are recorded.

Attitude : Conventional AIs remain neutral and objective; Sienna holds viewpoints and a stance.

Reference

Further details and examples are available on Sienna’s blog: https://sienna-blog-clone.vercel.app.