Dev shares 6-layer prompt pipeline with dual FAISS indexes to keep AI personas stable across models

OrionForgeEcosystem · reddit · 2026-11-14

A solo developer open-sourced SoulScript Engine, a system that keeps an AI character's personality, values and memories stable whether it runs on DeepSeek, Claude, GPT, Gemini or local Ollama.

Core idea: separate identity from memory

Identity index (read-only at runtime): a structured markdown 'Soul Script' per character with core values, boundaries, personality design and symbolic memories. Only the creator can edit it; the model can't rewrite who it is mid-chat.
Memory vault (read-write): per-user vault.jsonl with its own FAISS index. The agent writes memories via [MEMORYSAVE] tags, with a 13-action memory tool and PII guard.

Six-layer prompt pipeline: base system prompt → Soul Script retrieval → always-on injected knowledge → memory vault retrieval → tools (permission-checked per agent YAML) → conversation history capped at 30k characters. Identity comes before memory and history.

Why it's more than prompt engineering: identity is retrieved per turn rather than stuffed wholesale; it's write-protected at runtime; users/memories/agents are fully isolated with copy-on-write. Since identity and memory live in files and indexes you own, switching models mid-week preserves values, boundaries and memories.

Honest limitations: weaker models still drift in tone when their context maxes out. Repo runs locally via docker compose or uvicorn.

Original post →

Dev shares 6-layer prompt pipeline with dual FAISS indexes to keep AI personas stable across models

More from coding & agent