Project Chimera — Evolution Roadmap (AGI + Dual-Mode Architecture)

[LarytheLord]

Project Chimera — Evolution Roadmap (AGI Enhancement + Dual-Mode)

This discussion tracks all planned enhancements for Chimera, informed by a comprehensive review of the latest AGI papers, open-source repos, and HuggingFace models (March 2026).

Key constraint: $0 budget. All implementations use free-tier APIs, local CPU models, or open-source tools.

---

Architecture: Dual-Mode

Chimera must work as both:

1. Standalone AGI research project — CLI agent with web search, file system, consciousness
2. KM connector — therapy backend for Knight Medicare via chimera-bridge/

See Issue #15 for the package restructure plan.

---

Enhancement Tiers

Tier 1 — Quick wins (<1 day each)

Issue	Enhancement	Impact
#16	Local emotion detection — HF distilroberta model, 7 Ekman emotions, CPU	Saves 40% Gemini calls, enables mood tracking
#17	Reflexion self-critique — agent evaluates own response before sending	Catches bad therapy advice, ~15% extra API cost
#17	Constitutional AI guardrails — enforce therapy safety principles	Never suggests stopping medication, always provides crisis lines
—	Emotion trajectory — store per-message emotions in LanceDB	Mood trend visualization, better stuck-pattern detection
—	Memory importance scoring — weight recall by crisis/breakthrough/significance	Better context retrieval, more relevant responses

Tier 2 — Significant (1-3 days each)

Issue	Enhancement	Impact
#18	Local LLM fallback — SmolLM2-360M GGUF on CPU	Zero-downtime, works offline
—	Multi-agent inner council — split into Assessment/Intervention/Empathy/Safety agents	Better therapy quality via specialization
#15	Standalone CLI — python -m chimera.main	Chimera works independently for AGI research

Tier 3 — Advanced research (1-2 weeks each)

Enhancement	Impact
Cognitive state machine — SOAR/ACT-R inspired architecture	Proper cognitive cycle: PERCEIVE→ATTEND→RETRIEVE→DELIBERATE→REFLECT→ACT→LEARN
DSPy prompt optimization — auto-tune Gemini prompts	Prompts improve over time without manual tuning
RLHF with therapist feedback — psychologist rates AI responses	Continuous quality improvement via preference learning
Introspective state compression — compress cognitive states into personality traits	Efficient long-term patient models

---

Key Research Findings

Models we'll use (all CPU, free):

• Emotion: j-hartmann/emotion-english-distilroberta-base (7 labels, 330MB)
• Embeddings: sentence-transformers/all-MiniLM-L6-v2 (already using, 80MB)
• Local LLM: SmolLM2-360M-Instruct GGUF Q4 (220MB, fast on CPU)
• Reward model: distilbert-base-uncased (already in RLHF module)

Patterns we'll adopt:

• Reflexion (Shinn et al., 2023) — self-critique before responding
• Constitutional AI (Anthropic, 2022) — principle-based safety
• MemGPT/Letta tiered memory — working memory + archival LanceDB
• Emotion trajectory tracking — per-message emotion labels stored with episodic memory

What we're NOT doing:

• No GPU requirements (everything CPU)
• No paid APIs (Gemini free tier only)
• No heavyweight frameworks (no LangChain/LangGraph — too many deps)
• No local model training (inference only for HF models)

---

Current issues

#	Title	Assignee	Priority
#9	Fix chimera/__init__.py + shims	Prit	🔴 Blocker
#15	Phase A — Standalone CLI + restructure	Prit	🔴 High
#16	Local emotion detection (HF model)	Prit	🟠 High
#17	Reflexion + Constitutional AI	Prit	🟠 High
#18	Local LLM fallback (SmolLM2 GGUF)	Prit	🟡 Medium

Priority order: #9 → #15 → #16 → #17 → #18

Full plan: docs/CHIMERA-EVOLUTION-PLAN.md in the KM repo.

[LarytheLord]

Update: Research-backed enhancements added

After comprehensive AGI research (arXiv, GitHub, HuggingFace — 285+ papers reviewed), added two high-impact issues:

#20 — ACT-R Memory Decay + Temporal Validity

Inspired by: engram-ai/engram (ACT-R activation), Zep/Graphiti (temporal knowledge)

• Memories decay naturally unless reinforced (psychologically realistic)
• Old patient concerns get invalidated when patient updates ("I'm sleeping better now")
• Zero API cost for decay (pure math), minimal cost for temporal checks

#21 — Feed Consciousness Back Into Prompts

Inspired by: GödelOS (strange-loop), CoALA (metacognitive controller)

• Narcissus already records patterns and metrics — but they never enter the prompt
• Agent will be self-aware of its own therapeutic patterns and adjust
• Prevents therapeutic ruts (same technique every session)
• Zero extra API cost — data is already computed

Updated priority order for Prit:

#9 → #15 → #16 → #21 → #17 → #20 → #18

(#21 moved up because it's zero-cost and high-impact — just adding existing data to the prompt)

Full research findings

Saved in docs/CHIMERA-EVOLUTION-PLAN.md in the KM repo. Key repos discovered:

• plastic-labs/honcho (667★) — "dreaming" for patient profiles
• engram-ai/engram (26★) — ACT-R memory decay
• tao-hpu/cognitive-workspace (95★) — three-tier memory
• IAAR-Shanghai/Awesome-AI-Memory (553★) — 285+ memory papers