Core Concepts
Adaptive Retrieval
A 7-layer self-improving system that continuously refines retrieval quality.
Overview
MetaMemory's adaptive retrieval system is a 7-layer stack that continuously learns which retrieval strategies work best for different query types. It starts with no assumptions and improves with every search.
Layer 1: Meta-Memory Rules
A library of 50+ LLM-discovered patterns that map query characteristics to effective retrieval strategies. For example: “temporal queries about recent events perform better with higher recency decay weights.” Rules require a minimum confidence of 0.6 to be applied.
Layer 2: Thompson Sampling
A Beta-Bernoulli multi-armed bandit that balances exploration and exploitation. Each strategy arm maintains a Beta(α, β) posterior distribution, starting from a uniform Beta(1,1) prior. On each query, the system samples from each arm's posterior and selects the strategy with the highest sample.
Layer 3: UCB Selection
Upper Confidence Bound selection provides a complementary exploration signal. It selects the arm that maximizes the mean reward plus an exploration bonus proportional to uncertainty, ensuring under-explored strategies get tried.
Layer 4: Gradient Boosting
After 100+ retrieval samples have accumulated, a gradient boosting model (50 decision stumps, learning rate η=0.1) activates. It predicts retrieval effectiveness based on query features, context, and historical performance.
Layer 5: Ensemble Orchestration
Combines all strategy signals using weighted averaging:
| Signal | Weight |
|---|---|
| ML (Gradient Boosting) | 0.4 |
| Collaborative Filtering | 0.3 |
| MAB (Thompson/UCB) | 0.2 |
| Meta-Memory Rules | 0.1 |
Layer 6: Bayesian Parameter Optimization
Optimizes 7 continuous parameters (similarity threshold, decay rate, channel weights, etc.) using Bayesian optimization. Targets the 75th percentile of retrieval quality rather than the mean, ensuring consistently good results rather than high-variance performance.
Layer 7: Online Drift Detection
Monitors retrieval quality in real-time. If average quality drops by 10% or more from the rolling baseline, the system triggers an automatic retrain of the gradient boosting model and resets the Bayesian optimizer, preventing degradation from distributional shift.