Persona Core implements Closed Persona Agents: AI personas with sealed identities that operate within strict containment boundaries rather than an open system prompt. Every chat turn runs through a multi-stage pipeline (classify → enforce → assemble → gate → generate → postcheck → commit) that enforces identity isolation, dynamic consent, and summary-only memory, no raw transcript recall, ever. A separate operator-facing dashboard proxies to the chat API and exposes governance controls end users never see: persona lifecycle management, growth proposal review/approval, goal discovery, and self-model synthesis. Both processes share one SQLite database and are provider-agnostic across Ollama, OpenAI, and Anthropic.
Screenshots from a live session. The persona's portrait is blurred in the header; conversation excerpts are chosen from a lighter moment in the chat rather than deeper personal territory the persona is also capable of holding.
Clean Start
Before a persona is added: no default character, no forced identity, session state builds from here.
Persistent, In-Character Memory
Maya, a long/short macro hedge fund manager persona, holding a consistent professional identity and recalling prior context ("Right, you mentioned that before...") across turns, backed by the summary-only memory pipeline, not raw transcript replay.
Voice, Relay, and Model Controls
Per-session dev controls: Kokoro-backed TTS for spoken responses, paired with browser-native speech recognition (STT) on the mic button for voice input, an auto-engage multi-persona relay mode, and the active model, provider-agnostic and running local inference at no per-token cost.
A fully separate process from the chat API, exposing the governance controls end users never see.
Growth Monitor & Self-Model Synthesis
Relationship pattern tracking, resource assessment, self-model synthesis, and growth proposals, every proposed behavior change routes through here for review before it can be applied.
Session Governance & Audit Chain
Operator-level session control across four states (locked, unlocked, cooldown, quarantined) plus a tamper-evident hash-chained audit log of every governance action taken.
