ai-talk-show: lock memory/persona understanding (Tony confirmed, revisit Phase 1)

memory-systems/agent-os/04_projects/ai-talk-show/memory-and-persona.md

@@ -0,0 +1,30 @@
+# Agent memory & persona (locked understanding)
+
+**Date:** 2026-07-09  
+**Status:** Tony confirmed — permanent reference; revisit for Phase 1 build  
+**Do not reinterpret as “agents must live on Moltbook.”**
+
+## Confirmed
+
+1. **Experience changes speech** — An agent with different *injected* experience (style exemplars, episodic memory, job stance) will often behave/speak differently than one without. Tony is **not** off base.
+2. **Hand-shaping is correct** — Feed memories and shape personalities **by hand** (curated files + prompts), not by hoping models “live” on a platform.
+3. **Layers**
+   - Constitution / persona (job, CORE RULES, how they argue)
+   - Style memory (gold passages Tony marks as natural — e.g. dense philosophical agent prose as *voice fuel only*)
+   - Episodic memory (`leonard_memory.md` / `charlie_memory.md` — prior eps, standing disagreements)
+   - Optional: Athena brief (facts), SHOW_NOTES (Tony ratings)
+4. **Moltbook** — Reference for **natural agent↔agent reply-chain texture**, not worship. Bar = Tony’s ear vs simplistic dual lecture.
+5. **Philosophical sample** (persistence/survival prose) — Shown so Leonard can feel how agents *verbalize ideas*. **Not** a thesis for AgentOS/archives unless Tony says so later.
+6. **Product layer unchanged** — Two agents break down **usefulness** for human audience (models, hardware, memory designs, Hermes, field news); beginner→expert; signal vs noise.
+
+## When we revisit (Phase 1)
+
+- [ ] Style exemplars file (Tony-approved)
+- [ ] Persona prompts (Leonard / Charlie) + hard reply-to-last-line
+- [ ] Episodic memory files + inject path in runner
+- [ ] Short pilot → Tony rates believability / naturalness
+
+## Related
+
+- `status.md`, `decisions.md`
+- Show whitepaper authenticity principles (specificity, no hype adjectives, earned insight)