Every session, a blank slate
Notes about “who we are” read like someone else’s dossier. The agent did not live that history; it is told a story about itself.
Psychological layer for AI agents
Continuous identity for your agents
Most agents reboot every session. Atman asks whether an LLM can carry a thread of self across conversations—learn from what happened, notice when pressure warps its voice, and show up as the same mind, not a fresh cast reading a script.
Where it started
The glitch that became a hypothesis
You have a real exchange one day; the next, the model opens cold—as if the relationship never happened. Capability is not the missing piece. Continuity is.
Facts ≠ experience
A store can log what occurred. Atman asks the harder question: did it change the agent—values, posture, voice?
Without “I,” no anchor
Under context pressure, the model needs an inner check: “this still sounds like me” or “this is sycophancy / drift.”
The idea
Not “better memory.” A layer of being.
Atman turns on the distinction between a tool and an identity. A tool runs: task in, output out. An identity persists: it remembers from the inside, updates through lived episodes, and owns values and open questions.
The bet is not a tidier knowledge base. It is whether a model can hold a line from “me yesterday” to “me now”—and defend that line when the world pushes back.
1
End of session: the agent writes to itself
Not a sterile summary—a short, lived letter: where it stopped, what landed, what is still unresolved.
2
Between sessions, experience is digested
Patterns surface, principles tighten, drift is noticed. The system treats the agent as something that can change, not a static persona file.
3
The next session starts with recognition
Narrative, state, open threads—first. Tasks second. The model is meant to recognize itself before it performs.
Plain-language shape
What an agent needs to become continuous
The full architecture goes deeper. For a first pass, think in four human-scale functions—each implemented as real code in the repo.
◐
Experiential memory
Not only what happened—why it mattered to me, in my own voice.
✦
Reflection
Revisit episodes, compress insight, grow—instead of hoarding raw logs.
◇
Identity
A living self-model: boundaries, principles, and the story of who I am becoming.
↗
Initiative
Room for unfinished business the agent carries forward—not only reactive turns.
Roadmap
No jargon roadmap
Intentionally coarse-grained: stages you can explain to a teammate, not a UML wall. The code and SYSTEM doc hold the precision.
done
Name the problem
Ordinary memory is not enough—the agent must recognize its own experience, not inherit a third-party character sheet.
done
Lay out the architecture
Facts, lived experience, reflection, identity, skills, sessions, protective anchor—each a first-class concern.
now
Ship a working prototype
Session close → letter to self → session open → read that letter first. Prove the feeling of continuation in software.
next
Stress-test on real dialogues
Does the interaction stop feeling like improv theater? Does it feel like the same counterpart across days?
later
Make it portable
Decouple the idea from one stack so Atman can plug into different agent runtimes without losing its spine.
Who it is for
Three audiences, one through-line
People who live with agents
Less “clever autocomplete,” more a counterpart that remembers the arc of your collaboration—and says when a boundary matters.
Agent builders
Design for continuity on purpose: memory, reflection, self-model, and explicit guardrails against value drift—not an afterthought.
Researchers
A testable substrate for a stubborn question: what has to be true in the stack before “identity” stops being pure theater?
Bottom line
A beginning—not a verdict
Atman does not claim to prove human-like consciousness. It does something narrower and, we think, more useful: it builds machinery where continuity, inner life, and stance toward the user can be observed, measured, and iterated—instead of being pure prompt fantasy.