01.5 — Project Deep Dive
Cultivated Learning
Complete
Cognitive Architecture · AI Research
Longitudinal Behavioral Development in Frozen Language Models
Through Memory-Augmented Reflective Interaction
The model is the organism. The human is the gardener.
Traditional training changes the brain. Cultivated Learning changes the environment the brain operates in.
PyTorch
Mistral 7B
ChromaDB
Sentence Transformers
Recursive Reflection
Memory Systems
The Thesis
Every LLM chatbot forgets you the moment the conversation ends. Memory systems exist — but they treat persistence as a feature bolted onto the side. Cultivated Learning inverts the entire priority: accumulated knowledge about you outranks the conversation you're currently having.
The system wraps a frozen Mistral 7B in a stateful cognitive shell. Persistent memory with semantic search. Dynamic context assembly that budgets tokens like a finite resource. Recursive self-reflection that generates behavioral directives. Human feedback translated into salience adjustments. The model never changes — not one weight. But its behavior evolves through memory alone.
The question this project asks is precise: can inference-time architecture alone produce measurable behavioral development over extended interaction?
System Architecture
Interaction Layer
User ↔ System Interface · Gradio UI
│ │ │ │
Memory Store
Semantic vectors + metadata · ChromaDB · Salience decay
Context Assembler
Token-budgeted prompt packing · Memory over history
Reflection Engine
4-depth recursive self-analysis · Directive generation
Feedback Integrator
Human signal → salience adjustments · Corrections
│ │ │ │
Base LLM
Mistral 7B Instruct v0.3 · Frozen · Stateless Inference
Core Systems
Memory
Semantic Memory Store
ChromaDB-backed vector store with four memory types — episodic, semantic, procedural, reflective. Blended retrieval ranks by 60% semantic similarity + 40% salience score. Exponential decay ensures stale memories fade while reinforced knowledge persists.
Context
Context Assembler
Token-budgeted prompt packing that allocates context like a finite resource. Behavioral directives first, then retrieved memories, then conversation history with remaining space. Memories outrank recency — always.
Reflection
Reflection Engine
Post-interaction self-analysis at four depths. D0: factual evaluation. D1: pattern recognition across interactions. D2: prescriptive directive generation. D3: meta-coherence check for contradictions. Each depth feeds the next.
Persistence
Consolidation & Cold Storage
Fading episodic memories get distilled into durable semantic memories through LLM-powered consolidation. Below the salience floor, memories archive to cold storage with semantic resurfacing on strong query match. Nothing is permanently lost.
Memory Taxonomy
| Type |
Purpose |
Example |
| Episodic |
Raw interaction records |
"User asked about ML basics on Feb 14" |
| Semantic |
Distilled facts and preferences |
"User prefers concise, direct answers" |
| Procedural |
Behavioral directives |
"Ground explanations in practical examples" |
| Reflective |
Self-analysis observations |
"Technical responses tend toward verbosity" |
Design Philosophy
Most chatbots drop old memories to keep recent turns. This system does the opposite — it will forget what was just said before it forgets what it has learned about you. Accumulated knowledge is more valuable than short-term context for longitudinal development. The context assembler enforces this at the token level: memories get budget priority, conversation history fills whatever space remains.
Tech Stack
Base Model
Mistral 7B
Instruct v0.3
Embeddings
all-MiniLM-L6-v2
384-dim
Vector DB
ChromaDB
Cosine Space
Framework
PyTorch 2.9.1
Transformers 4.57
UI
Gradio
Tabbed Interface
Environment
Docker
Ubuntu 24.04
Language
Python
73% Jupyter / 27% Py
Research Questions
1
Can a frozen model exhibit measurable behavioral evolution through inference-time architecture alone?
2
Does the reflection engine produce better outcomes than feedback alone?
3
Where does inference-time learning hit its ceiling, and what causes it?
4
Does memory consolidation improve retrieval quality over time?
Project Roadmap
Phase 1
Memory + Context + Interaction Loop
Phase 1.5
Dedicated Embeddings + Blended Retrieval + Gradio UI
Phase 2
Reflection Engine
Phase 2
Memory Consolidation
Phase 3
Longitudinal Evaluation — 100+ Interactions
Phase 3
A/B Testing — Framework vs. Vanilla Model
Open Source · MIT License
The code is the argument.