feat: Implement episodic memory layer and experiment gating with LLM resolution

memory.md

@@ -0,0 +1,46 @@
+# Persistent Memory Layer for Autoresearch
+
+Built a persistent memory layer for @karpathy's `autoresearch` — an autonomous ML research agent that runs hundreds of experiments overnight.
+
+## The Problem
+The agent rediscovers the same dead ends every run. It has no memory of what it already tried.
+
+## The Fix
+- **Persistent Storage**: Every experiment (hyperparameters + `val_bpb`) gets stored in SQLite with a confidence score.
+- **Memory Gate**: Before proposing the next experiment, the agent queries memory first — that's the gate.
+- **Normalization**: Hyperparameters are std-normalized before comparison, so `DEPTH=8` and `LR=0.04` don't distort each other's scale.
+- **Retrieval**: Cosine similarity finds the closest past experiments in normalized hyperparameter space.
+- **Confidence Updates**: High similarity + same verdict → confidence goes up, no LLM needed.
+- **Conflict Resolution**: High similarity + opposite verdict → single LLM-as-judge call to resolve the contradiction.
+- **Ground Truth**: Resolved verdict gets written back as ground truth.
+
+## Flow
+Agent proposes experiment
+        ↓
+Query memory (should_run_experiment)
+        ↓
+Normalize hyperparameters (std scaling)
+        ↓
+Cosine similarity → find closest past experiments
+        ↓
+High similarity + same verdict   → update confidence
+High similarity + opposite verdict → LLM resolves
+        ↓
+Verdict written back as ground truth
+
+One expensive operation (LLM call) only when the math can't decide. Everything else is pure geometry.
+
+
+## Quick Start
+```bash
+pip install pydantic requests
+python test.py
+```
+To enable LLM conflict resolution:
+```bash
+export OPENAI_API_KEY=your_key
+python test.py
+```
+
+## Inspiration
+The architecture is directly inspired by CoALA (Sumers et al. 2023) — episodic memory for autonomous research agents.