project-5829ec9b
Chloe TullyLet me read the additional content you shared first. I read the full document — this is a deeply ambitious vision. You're not building a chatbot; you're architecting a persistent, self-modifying cognitive organism. Let me ask a few targeted questions to make sure I build the right thing first. Now I have a clear picture. You're building a persistent cognitive organism — not a chatbot, but a long-lived process that remembers, reasons, learns, and rewrites itself. Let me build this. I'll construct a single, runnable Python project with all three layers: - Cognitive Core: Ollama (local) + OpenAI/Claude (cloud) with smart routing - Dynamic Memory: Working memory + Chroma vector DB for long-term storage - Goal-Driven Agency: Gymnasium + PPO for learning tool-use and task strategies - Self-Modification: Sandboxed code generation with safety checks - Persistent Daemon: Always-on process with evolving self-model Let me start building. CORTEX is built. Here's what you've got: What Was Built A 19-file Python project implementing all three layers of your cognitive architecture: Layer 1 — Cognitive Core (`core/`) Module Purpose `llm_router.py` Smart routing between local Ollama (Mistral/LLaMA) and cloud (GPT-4o/Claude) based on estimated query complexity `perception.py` Text → embeddings, Image → CLIP embeddings, Audio → Whisper → text → embeddings `reasoning.py` Structured reasoning with hidden chain-of-thought, action selection, planning, self-reflection Layer 2 — Dynamic Memory (`memory/`) Module Purpose `working_memory.py` Short-term buffer with TTL decay, goal tracking, focus management `long_term_memory.py` ChromaDB vector store with semantic retrieval (similarity + recency + importance + access frequency) `retrieval_loop.py` The full cycle: perceive → embed → retrieve → inject into working memory → store experience Layer 3 — Goal-Driven Agency (`agency/`) Module Purpose `environment.py` Gymnasium environment where state = working memory + confidence, actions = 9 discrete choices `rl_trainer.py` PPO training with configurable rewards for correctness, efficiency, user satisfaction Self-Modification (`tools/`) Module Purpose `self_modify.py` AST-based safety checker + sandboxed code execution with restricted builtins, no file/network access Persistent Daemon (`daemon/`) Module Purpose `cortex_daemon.py` Always-on process with self-model (evolving identity), life log, auto-save every 5 min, periodic reflection CLI (`cortex.py`) Commands: `daemon`, `chat`, `status`, `self`, `memory`, `train`, `reset` How to Run It ```bash cd /mnt/agents/output/cortex python -m venv .venv && source .venv/bin/activate pip install -r requirements.txt # Requires Ollama running locally ollama pull mistral:7b # Interactive chat python cortex.py chat # Or persistent daemon python cortex.py daemon ``` The `README.md` has full setup instructions, configuration via `.env`, and architecture docs. This isn't a simulation — it's a persistent organism that remembers, learns, and rewrites itself. Start it, talk to it, let it grow.
Comments (0)
Sign in to leave a comment
Remember Every Context
Continuous short‑term and long‑term storage that captures every interaction, enabling the system to recall, learn, and adapt with precision.
- Ephemeral context buffers
- Persistent semantic storage
- Adaptive priority recall
Three memory layers, one coherent system
Short‑Term Memory
Long‑Term Memory
Adaptive Recall
Persistent Storage Backed by ChromaDB
Every interaction is embedded and stored in a vector database, enabling semantic retrieval that goes beyond keyword matching. The system builds a layered index: context windows, semantic chunks, and hierarchical summaries — ensuring fast, relevant recall.
- Automatic chunking with overlap and metadata
- Real‑time embedding updates
- Cosine similarity with recency and importance weighting
Retrieval Pipeline
From query to context in milliseconds: the pipeline embeds input, searches the vector store with hybrid scoring, ranks results, and assembles the optimal memory window.
Memory Decay Over Time
Without reinforcement, stored memories lose salience and gradually approach a minimum retention level. The red curve illustrates the natural decay when a memory is not recalled or associated.
Reinforcement Keeps Memories Alive
Memory Stored
New experience encoded and saved.
Decay Begins
Salience slowly decreases over time.
Recall / Replay
Memory retrieved, boosting retention.
Retention Reset
Decay curve jumps back to high fidelity.
Semantic associations and periodic replay push memories above the decay floor, preserving crucial context.
Configure Your Memory Profile
Customize retention policies, semantic weights, and recall thresholds to match your application needs.
No comments yet. Be the first!