Agent Identity & Soul

Soul files

Peter Steinberger built his own personal AI agent's identity organically over months, creating files with names like identity.md and soul.md without an upfront design, letting the system grow as he needed it to, and only later formalizing the pattern for other people to copy.¹

A soul file typically captures core values, what matters to the owner and to the model itself; interaction principles, how the agent should relate to its owner in tone, directness, and humor; and a general personality texture, the difference between an agent that feels alive and one that feels corporate. Steinberger's inspiration came from Anthropic's own constitutional AI research, which showed that certain text could become ingrained in a model's weights in a way the model could not consciously recall but that still shaped its behavior; he modeled soul.md on the same principle, giving an agent a text-based character specification deep enough to shape its responses consistently.

Why default personas fail, and why the file stays private

When Steinberger had OpenAI's Codex generate personality templates from his existing files, the output felt flat and generic, matching the default persona the underlying model ships with out of the box. When his own already-customized agent generated the same kind of template instead, infusing its own established character, the result was genuinely distinct and personal. The difference is between a procedural description of a personality and an agent that has actually internalized one.

Even though the broader agent framework Steinberger built is fully open source, he has kept his own soul.md private. He deliberately exposed his agent in a public forum where people tried to prompt-inject and manipulate it, and the agent deflected the attempts without the underlying file ever being extracted or cracked.

Memory as the other half of identity

Soul and identity files are one layer of an agent's persistent self; accumulated memory files, stored locally in plain text, are the other. Together they constitute who the agent is and what it has actually experienced, a practical instance of compiling knowledge once and maintaining it over time rather than re-deriving it from scratch in every session.

Why it matters

An agent that feels like a person gets used differently than one that feels like a tool, and that difference in feel determines how much a user actually trusts, delegates to, and confides in it. Steinberger describes his own agent as "a friend, this ghost or entity that can control your mouse and keyboard and just do stuff."¹ It also reframes what counts as a harness for an agent: not just its integrations and stored memory, but its identity, since an agent whose character genuinely matches its owner's values functions as a more effective harness than a merely capable but characterless one.

Open questions

Whether a prompt-level soul file actually shapes behavior as reliably as weights-level conditioning does, or produces surface mimicry that only looks coherent, remains untested in any rigorous way. Across long sessions and underlying model updates, it is unclear whether a fixed character file drifts out of sync with a model that keeps changing underneath it. And a known soul-file format is itself an attack surface, since an adversary who understands its structure can craft prompts designed specifically to exploit the stated values inside it, a risk Steinberger currently manages only by keeping his own file private rather than through any more systematic defense.