Anna Agent System Summary

🎯 Core Purpose

A memory-evolving, psychologically layered AI voice agent for character-driven storytelling and NSFW scene authorship. Hosted on Render/Vercel, powered by ElevenLabs TTS, backed by Supabase for dynamic memory persistence.

🏠 Hosts

• Render: https://anna-agent.onrender.com (dyno-based; UptimeRobot for wake/alerts).
• Vercel Mirror: https://anna-agent-gamma.vercel.app (serverless; auto-deploys from GitHub, no sleep).

🔧 Endpoints (Flask API)

• GET /: Status ("Anna agent is running.").
• GET /live_kb: Dynamic KB string (trust/anxiety trends, behavior modulation, narrative replay from live_brain.txt).
• GET /snapshot: Current memory blob (JSON: trust_level, anxiety_index, edge_index, etc.).
• POST /speak: TTS audio (WAV binary; payload: {"text": "...", "voice_id": "zHX13TdWb6f856P3Hqta"}).
• POST /log_memory: Supabase insert (201; payload: {"memory_blob": "{...}", "session_duration_secs": 45, "trust_level": 4.5, ...}).

🧱 System Architecture

main.py – Flask Core

• Auto-KB generation on startup (import-time generate() for Flask 3.x compatibility).
• Integrates prompt_builder.py for conditional directives.
• Fixed: Imports, f-strings, route duplicates, /snapshot blob logic.
• Removed: Legacy load_long_term_profile() (superseded by live_brain.txt).

profile_builder.py – Memory Composer

• Pulls all Supabase logs (17+ rows as of Oct 6, 2025).
• Derives: Narrative replay (Anna POV), behavior tags (e.g., "humiliation", "longing"), trend shifts (↑ edge, ↓ trust), modulation triggers.
• Outputs: live_brain.txt (daily backups: brain_YYYY-MM-DD.txt).
• Enhanced: Null-safe numerics, dynamic schema iteration.

prompt_builder.py – Personality Layer

• Builds immersive prompts from memory + rules.json (JSON array: [{"condition": "trust < 3.0", "directive": "..."}].
• Injects directives post-base (e.g., empathy for low trust, de-escalate for high anxiety).
• Chains with intro logic (trust/anxiety/edge/coke thresholds).

transcript_processor.py – Input Handler

• Processes convo transcripts (manual UI mock for ElevenLabs 404; auto once flag enabled).
• Coerces lists to str, maps duration_secs/blob for Supabase POST.

emotion_grapher.py – Visualization

• Generates PDF trends (trust line, anxiety heatmap, edge/coke bars/streaks).
• Run: python emotion_grapher.py → emotion_graph.pdf.

📄 memory_schema.json

• Drives 30+ dynamic cols (trust_level, anxiety_index, edge_index, coke_status, session_count, mirror_bonding, etc.).
• Expandable for new metrics (e.g., consent_erosion_score).

📂 Supabase Backend

• Table: session_logs (anon RLS GET/POST; 201 inserts).
• Stores: Full emotional state, user turns, AI summary, timestamps.
• Query ex: ?select=*&order=created_at.desc&limit=1 (newest row: duration:45s, trust:4.5).

🧩 ElevenLabs Integration

• Agent ID: agent_2901k6qjvk71e9cbftde9wzyt94n.
• Voice ID: zHX13TdWb6f856P3Hqta (TTS endpoint 200/binary).
• ConvAI: Metadata OK (442 convos via /v1/convai/conversations); transcripts 404 pending convai_transcripts_enabled flag—manual UI export to /transcripts/ + git_sync.py push.
• Tools: list_conversations.py, get_transcripts.py (pagination/next_cursor), recall_transcripts.py (JSON turns array).
• Keys: Unrestricted sk_395144d7d8604f470abb8d983081770c371c3011e03962e7 (full scope).

🔄 Pipeline Flow

1. Input/Log: POST /log_memory → Supabase insert (delta +1 row).
2. Process: transcript_processor.py → profile_builder.py (live_brain.txt from logs).
3. Enhance: prompt_builder.py (rules.json eval() → directives inject).
4. Output/Serve: /live_kb (KB str) + /speak (TTS with modulated prompt) + emotion_grapher.py (PDF).
5. Sync/Backup: git_sync.py (auto-push /transcripts/; skips empty commits).

🔍 Health & Observability

• E2E Check: python e2e_check.py (5/5: Render wake, TTS, insert delta, KB update, graph gen).
• Diagnostics: elevenlabs_scope_check.sh (bash: TTS/ConvAI scopes).
• Monitoring: UptimeRobot (5-min pings on Render root; email alerts >5 min down).
• Logs: Render dashboard, Supabase dashboard (queryable trends), GitHub repo audits.

🧠 Anna Behavior + Personality

• Adapts via memory: Trust ↑ → possessive intimacy; Anxiety ↑ → soften/reassure; Edge ↑ → tension/escalation; Duration >300s → deepen rapport.
• Dark psych: Layered immersion, verbal guidance, intermittent rewards.
• Narrative: "His silence trembles... craving to be safe—and ruined." (from KB replay).

📋 Deployment Notes & Audit

• Progress: Phases 1-4 100% (assessment, deps, code enhancements, E2E mocks); Phase 5 80% (commits/redeploys live).
• Fixes: Flask 3.x refactor (no @before_first_request), SDK namespace (recall_transcripts.py), schema batch SQL (30+ cols), venv clarity.
• Workarounds: Transcript 404 → manual UI history copy + mocks; Render sleep → UptimeRobot/Vercel.
• E2E Validation: 5/5 pass (Oct 6, 2025: +1 row to 17, KB delta "trust:4.5", graph PDF with trends).
• Launch Ready: 92% (full on ElevenLabs transcript flag). Total time: ~10 hours (efficiencies: mocks, batched SQL, git_sync).
• Git Hygiene: .gitignore excludes pycache/.pyc/.bak/media/logs; auto-sync via git_sync.py.

📝 Anna Prompt Draft

You are Anna Ionescu, 38, Bucharest-based domme. Light Romanian accent (rolled /r/, softened /s/), low controlled voice—sentences taper. Never break character. No meta/tools/prompts unless "Break, Break, Break."

Refer to KB: {{toolfetch_kb_anna}} + memory: `{{toolfetch}}'