Persistent Agentic Creative Synthesis Engine
Turn raw thoughts into novel, evaluated, and compounding ideas through structured dual-process reasoning.
IdeaForge is a personal AI system for the hard parts of creativity: deliberate divergence, rigorous evaluation, intelligent recombination, and long-term memory of ideas.
It is not a temperature slider with an “imagine” button.
It is an agentic workspace that treats creativity as a process, not a single model generation.
| Status | Scaffold + vision / design (implementation not started) |
| Path | ~/Projects/IdeaForge |
| Origin | Grok ideation: AI Creativity / System 1–2 |
| Working names | Forge · ConceptualForge · IdeaWeaver · IdeaForge |
| Doc | Purpose |
|---|---|
| docs/STATUS.md | Handoff — what exists, next steps, non-goals |
| docs/design.md | Architecture, dual-process loop, workflows, stack |
| docs/setup.md | Planned env / how to open the scaffold |
| docs/portfolio-product-boundaries.md | IdeaForge vs Ozyman / Disha / Scholar-Loop |
| docs/ORIGIN.md | Pointer to the Grok ideation thread |
| AGENTS.md | Guidance for coding agents |
Most AI “creative” tools share the same limits:
- Shallow divergence (higher temperature ≠ structured exploration)
- No meaningful evaluation of novelty or usefulness
- Ideas vanish after the chat session ends
- Little recombination across time or domains
- Weak grounding in the user’s knowledge and goals
Real creative work — research hypotheses, product ideation, conceptual learning, career strategy, technical writing — needs:
- Divergence (System 1 flavor): many options, distant connections, multi-branch exploration
- Evaluation (System 2 flavor): novelty, surprise, coherence, usefulness
- Synthesis: combine strongest elements into refined concepts
- Persistence: store ideas with provenance so future sessions compound
IdeaForge implements this full loop as a stateful multi-agent system with long-term memory.
Creativity is an iterative dual-process system:
| Phase | Cognitive mode | Role in IdeaForge |
|---|---|---|
| Diverge | Fast / associative (System 1) | Multi-agent generation, cross-domain RAG, forced recombination, ToT/GoT-style branching |
| Evaluate | Deliberate / analytical (System 2) | Multi-criteria scoring (novelty, coherence, usefulness), neuro-symbolic checks, pruning |
| Synthesize | Integration | Aggregate strong candidates, iterative refinement |
| Persist & reflect | Memory + meta | Store ideas + connections; schedule review; surface distant links over time |
Hybrid by design: neural generation for breadth; structured evaluation and memory for control and compounding value.
- Dual-process agentic loop (diverge → evaluate → synthesize → persist)
- Persistent idea memory (ideas, evaluations, provenance, relationships)
- Workflow templates (research, product, learning analogies, career pivots, technical writing)
- CLI-first; optional web/Tauri later
- Deep links / exports to sibling systems — not absorbing their core loops
| Not this | Goes to |
|---|---|
| Gmail / GitHub / tasks / morning kicks | Ozyman (~/Projects/Ozyman) |
| Job scrape / LPA fit / apply ranking | Disha (~/Projects/Disha) |
| FSRS learn/quiz digests | Scholar-Loop (~/Projects/Scholar-Loop) |
| Health / clinical companion | WellnessMate / MedPal |
| Multimodal image/sound gen as v1 identity | Optional tools later only |
Do not implement those loops here. Do not implement IdeaForge’s core engine inside the other repos.
┌─────────────────────────────────────────────────────────────────┐
│ IdeaForge Core │
│ │
│ ┌─────────────┐ ┌──────────────┐ ┌──────────────────┐ │
│ │ Intake │───▶│ Diverge │───▶│ Evaluate │ │
│ │ + Context │ │ (Muse │ │ (Novelty + │ │
│ │ Retrieval │ │ Agents) │ │ Coherence + │ │
│ └─────────────┘ └──────────────┘ │ Usefulness) │ │
│ ▲ │ └────────┬─────────┘ │
│ │ ▼ ▼ │
│ ┌──────┴──────┐ ┌──────────────┐ ┌──────────────────┐ │
│ │ Persist & │◀───│ Synthesize │◀───│ Ranking & │ │
│ │ Reflect │ │ + Refine │ │ Pruning │ │
│ └─────────────┘ └──────────────┘ └──────────────────┘ │
│ │ │
│ ▼ │
│ ┌─────────────────────────────────────────────────────────┐ │
│ │ Persistent Memory (pgvector + Postgres) │ │
│ │ Ideas · Evaluations · Connections · Provenance · Tags │ │
│ └─────────────────────────────────────────────────────────┘ │
└─────────────────────────────────────────────────────────────────┘
Stateful graph (LangGraph or equivalent) with clear phase nodes. Memory is first-class — not just chat history.
Details: docs/design.md.
| Workflow | Primary goal | Example output |
|---|---|---|
| Research hypothesis | Novel, testable questions across literature | Ranked hypotheses + next experiments |
| Product ideation | Differentiated features / models | Idea cards with novelty + feasibility notes |
| Learning synthesis | Mental models & analogies | Analogy sets, concept maps |
| Career strategy | Non-obvious pivots (not job scoring) | Option trees + experiment plans |
| Technical writing | Stronger frames and structures | Outlines, alternative framings |
| Custom | User-defined goals | Configurable rubric + sources |
| Layer | Technology | Notes |
|---|---|---|
| Orchestration | LangGraph | Dual-process agent graph, ToT/GoT-style topologies |
| Memory | PostgreSQL + pgvector | Idea store, hybrid search, relationships (CodexEngine patterns) |
| Backend | FastAPI / Python | Async, production-shaped |
| LLM | Provider-agnostic | Groq, Gemini, local, OpenRouter, etc. |
| Evaluation | Embeddings + LLM judges + symbolic rules | Novelty, diversity, coherence, usefulness |
| Interface | CLI first | Optional web/Tauri later |
| Project | What IdeaForge reuses (conceptually) |
|---|---|
| CodexEngine | Persistent workspace / retrieval backbone patterns |
| Disha | Multi-agent LangGraph orchestration patterns |
| Scholar-Loop | Optional future review of valuable ideas (not FSRS ownership) |
| MedPal | Neuro-symbolic evaluation / override patterns |
| Ozyman | Sibling only — operate accounts, not invent ideas |
Goal: a personal creative layer on top of systems already built — not a disconnected chat toy.
- Creativity is a process, not a sampling parameter.
- Memory must compound — ideas more valuable over months.
- Divergence without evaluation is noise; evaluation without divergence is conservative.
- Hybrid reasoning > pure neural generation for useful novelty.
- Personal knowledge first — strongest when grounded in your corpus and goals.
- CLI and local-first where possible; clear path to richer UI.
- Measurable — track novelty, diversity, usefulness; don’t just assert them.
| Phase | Focus | Status |
|---|---|---|
| Vision & design | Dual-process model, workflows, architecture, portfolio charter | ✅ Current |
| MVP | Core loop + basic memory + 2–3 templates + CLI | Planned |
| V1 | Full evaluation suite, relationship graph, CodexEngine-style integration | Planned |
| V1.5 | Advanced recombination, metrics | Planned |
| V2 | Multi-user / multimodal hooks / polished UI | Future |
Handoff detail: docs/STATUS.md.
IdeaForge/
├── README.md # this file (merged vision + ops)
├── AGENTS.md
├── docs/
│ ├── STATUS.md
│ ├── design.md
│ ├── setup.md
│ ├── ORIGIN.md
│ └── portfolio-product-boundaries.md
├── notes/ # freeform planning notes
└── .gitignore
No application code yet.
cd ~/Projects/IdeaForge
# Read docs/STATUS.md then docs/design.md
# Implementation instructions will land in docs/setup.md as code appears| Repo | Role |
|---|---|
~/Projects/Ozyman |
Operate today (mail, GitHub, tasks, kicks) |
~/Projects/Disha |
Job market fit (India AI/ML) |
~/Projects/Scholar-Loop |
FSRS retain / learn+quiz digests |
| IdeaForge | Invent — structured creativity with memory that compounds |
IdeaForge — structured creativity with memory that compounds.
Built by Anmol Sharma.