diff --git a/Dockerfile.mcp b/Dockerfile.mcp
index 22524111..c6c85d63 100644
--- a/Dockerfile.mcp
+++ b/Dockerfile.mcp
@@ -13,6 +13,10 @@ RUN pip install --no-cache-dir --upgrade mcp fastmcp qdrant-client fastembed
 # Create cache directory with proper permissions
 RUN mkdir -p /tmp/cache && chmod 755 /tmp/cache
 
+# Ensure rerank volumes are writable when containers run as non-root (uid 1000).
+RUN mkdir -p /tmp/rerank_events /tmp/rerank_weights \
+    && chmod 777 /tmp/rerank_events /tmp/rerank_weights
+
 # Bake scripts into image so server can run even when /work points elsewhere
 COPY scripts /app/scripts
 
diff --git a/Dockerfile.mcp-indexer b/Dockerfile.mcp-indexer
index 5cfe1286..9e7a6281 100644
--- a/Dockerfile.mcp-indexer
+++ b/Dockerfile.mcp-indexer
@@ -14,6 +14,12 @@ RUN apt-get update && apt-get install -y --no-install-recommends git ca-certific
 COPY requirements.txt /tmp/requirements.txt
 RUN pip install --no-cache-dir --upgrade -r /tmp/requirements.txt
 
+# Ensure rerank volumes are writable when containers run as non-root (uid 1000).
+# On first mount of an empty named volume, Docker copies directory contents (incl perms)
+# from the image into the volume.
+RUN mkdir -p /tmp/rerank_events /tmp/rerank_weights \
+    && chmod 777 /tmp/rerank_events /tmp/rerank_weights
+
 # Download reranker model and tokenizer during build
 # Cross-encoder for reranking (ms-marco-MiniLM) + BGE tokenizer for micro-chunking
 ARG RERANKER_ONNX_URL=https://huggingface.co/cross-encoder/ms-marco-MiniLM-L-6-v2/resolve/main/onnx/model.onnx
diff --git a/docker-compose.yml b/docker-compose.yml
index a8b26557..b77fc8d6 100644
--- a/docker-compose.yml
+++ b/docker-compose.yml
@@ -59,11 +59,18 @@ services:
       - TOOL_STORE_DESCRIPTION=${TOOL_STORE_DESCRIPTION}
       - TOOL_FIND_DESCRIPTION=${TOOL_FIND_DESCRIPTION}
       - FASTMCP_HEALTH_PORT=18000
+      # Learning reranker configuration
+      - RERANK_LEARNING=${RERANK_LEARNING:-1}
+      - RERANKER_WEIGHTS_DIR=/tmp/rerank_weights
+      - RERANK_EVENTS_DIR=/tmp/rerank_events
+      - RERANK_EVENTS_ENABLED=${RERANK_EVENTS_ENABLED:-1}
     ports:
       - "18000:18000"
       - "8000:8000"
     volumes:
       - workspace_pvc:/work:ro
+      - rerank_data:/tmp/rerank_weights:rw
+      - rerank_events:/tmp/rerank_events:rw
     networks:
       - dev-remote-network
 
@@ -120,15 +127,57 @@ services:
       - INDEX_UPSERT_BATCH=${INDEX_UPSERT_BATCH:-512}
       - INDEX_UPSERT_RETRIES=${INDEX_UPSERT_RETRIES:-5}
       - MAX_MICRO_CHUNKS_PER_FILE=${MAX_MICRO_CHUNKS_PER_FILE:-200}
+      # Learning reranker configuration
+      - RERANK_LEARNING=${RERANK_LEARNING:-1}
+      - RERANKER_WEIGHTS_DIR=/tmp/rerank_weights
+      - RERANK_EVENTS_DIR=/tmp/rerank_events
+      - RERANK_EVENTS_ENABLED=${RERANK_EVENTS_ENABLED:-1}
     ports:
       - "${FASTMCP_INDEXER_PORT:-8001}:8001"
       - "18001:18001"
     volumes:
       - workspace_pvc:/work:rw
       - codebase_pvc:/work/.codebase:rw
+      - rerank_data:/tmp/rerank_weights:rw
+      - rerank_events:/tmp/rerank_events:rw
     networks:
       - dev-remote-network
 
+  # Learning reranker worker - processes training events in background
+  learning_worker:
+    build:
+      context: .
+      dockerfile: Dockerfile.mcp-indexer
+    container_name: learning-worker-dev-remote
+    user: "1000:1000"
+    command: ["sh", "-c", "mkdir -p /tmp/huggingface/hub /tmp/huggingface/transformers /tmp/huggingface/fastembed && exec python /app/scripts/learning_reranker_worker.py --daemon"]
+    depends_on:
+      - qdrant
+      - mcp_indexer
+    env_file:
+      - .env
+    environment:
+      - QDRANT_URL=${QDRANT_URL}
+      - COLLECTION_NAME=${COLLECTION_NAME}
+      - HF_HOME=/tmp/huggingface
+      - HF_HUB_CACHE=/tmp/huggingface/hub
+      - TRANSFORMERS_CACHE=/tmp/huggingface/transformers
+      - FASTEMBED_CACHE_PATH=/tmp/huggingface/fastembed
+      - EMBEDDING_MODEL=${EMBEDDING_MODEL}
+      - EMBEDDING_PROVIDER=${EMBEDDING_PROVIDER}
+      - RERANK_EVENTS_DIR=/tmp/rerank_events
+      - RERANKER_WEIGHTS_DIR=/tmp/rerank_weights
+      - RERANK_LEARNING_BATCH_SIZE=${RERANK_LEARNING_BATCH_SIZE:-32}
+      - RERANK_LEARNING_POLL_INTERVAL=${RERANK_LEARNING_POLL_INTERVAL:-30}
+      - RERANK_LEARNING_RATE=${RERANK_LEARNING_RATE:-0.001}
+    volumes:
+      - workspace_pvc:/work:rw
+      - rerank_data:/tmp/rerank_weights:rw
+      - rerank_events:/tmp/rerank_events:rw
+    networks:
+      - dev-remote-network
+    restart: unless-stopped
+
   # MCP HTTP search service - same as base compose
   mcp_http:
     build:
@@ -171,11 +220,18 @@ services:
       - TOOL_STORE_DESCRIPTION=${TOOL_STORE_DESCRIPTION}
       - TOOL_FIND_DESCRIPTION=${TOOL_FIND_DESCRIPTION}
       - FASTMCP_HEALTH_PORT=18000
+      # Learning reranker configuration
+      - RERANK_LEARNING=${RERANK_LEARNING:-1}
+      - RERANKER_WEIGHTS_DIR=/tmp/rerank_weights
+      - RERANK_EVENTS_DIR=/tmp/rerank_events
+      - RERANK_EVENTS_ENABLED=${RERANK_EVENTS_ENABLED:-1}
     ports:
       - "${FASTMCP_HTTP_HEALTH_PORT:-18002}:18000"
       - "${FASTMCP_HTTP_PORT:-8002}:8000"
     volumes:
       - workspace_pvc:/work:ro
+      - rerank_data:/tmp/rerank_weights:rw
+      - rerank_events:/tmp/rerank_events:rw
     networks:
       - dev-remote-network
 
@@ -233,12 +289,19 @@ services:
       - INDEX_UPSERT_BATCH=${INDEX_UPSERT_BATCH:-512}
       - INDEX_UPSERT_RETRIES=${INDEX_UPSERT_RETRIES:-5}
       - MAX_MICRO_CHUNKS_PER_FILE=${MAX_MICRO_CHUNKS_PER_FILE:-200}
+      # Learning reranker configuration
+      - RERANK_LEARNING=${RERANK_LEARNING:-1}
+      - RERANKER_WEIGHTS_DIR=/tmp/rerank_weights
+      - RERANK_EVENTS_DIR=/tmp/rerank_events
+      - RERANK_EVENTS_ENABLED=${RERANK_EVENTS_ENABLED:-1}
     ports:
       - "${FASTMCP_INDEXER_HTTP_PORT:-8003}:8001"
       - "${FASTMCP_INDEXER_HTTP_HEALTH_PORT:-18003}:18001"
     volumes:
       - workspace_pvc:/work:rw
       - codebase_pvc:/work/.codebase:rw
+      - rerank_data:/tmp/rerank_weights:rw
+      - rerank_events:/tmp/rerank_events:rw
     networks:
       - dev-remote-network
 
@@ -494,6 +557,14 @@ volumes:
   qdrant_storage_dev_remote:
     driver: local
 
+  # Learning reranker weights storage (shared between indexer and worker)
+  rerank_data:
+    driver: local
+
+  # Learning reranker events storage (shared between indexer and worker)
+  rerank_events:
+    driver: local
+
 # Custom network for service discovery
 networks:
   dev-remote-network:
diff --git a/docs/ARCHITECTURE.md b/docs/ARCHITECTURE.md
index cc6edd43..b7549761 100644
--- a/docs/ARCHITECTURE.md
+++ b/docs/ARCHITECTURE.md
@@ -8,6 +8,7 @@
 - [Overview](#overview)
 - [Core Principles](#core-principles)
 - [System Architecture](#system-architecture)
+- [Learning Reranker System](#5-learning-reranker-system)
 - [Data Flow](#data-flow)
 - [ReFRAG Pipeline](#refrag-pipeline)
 
@@ -122,6 +123,110 @@ Context Engine is a production-ready MCP (Model Context Protocol) retrieval stac
 - **Local LLM Integration**: llama.cpp for offline expansion
 - **Caching**: Expanded query results cached for reuse
 
+### 5. Learning Reranker System (Optional)
+
+The Learning Reranker is an **optional** self-improving ranking system that learns from search patterns to provide increasingly relevant results over time. It is enabled by default but can be disabled via `RERANK_LEARNING=0` and `RERANK_EVENTS_ENABLED=0` environment variables. See [Configuration](CONFIGURATION.md#learning-reranker) for all options.
+
+#### Architecture Overview
+
+```
+┌─────────────────┐     ┌──────────────────┐     ┌─────────────────┐
+│  Search Query   │────►│  Hybrid Search   │────►│  TinyScorer     │
+│                 │     │  (initial rank)  │     │  (learned rank) │
+└─────────────────┘     └──────────────────┘     └─────────────────┘
+                                                         │
+                        ┌──────────────────┐             │
+                        │  Event Logger    │◄────────────┘
+                        │  (NDJSON files)  │
+                        └────────┬─────────┘
+                                 │
+                        ┌────────▼─────────┐
+                        │ Learning Worker  │
+                        │  (background)    │
+                        └────────┬─────────┘
+                                 │
+                        ┌────────▼─────────┐
+                        │  ONNX Teacher    │
+                        │ (cross-encoder)  │
+                        └────────┬─────────┘
+                                 │
+                        ┌────────▼─────────┐
+                        │  Weight Updates  │
+                        │  (.npz files)    │
+                        └──────────────────┘
+```
+
+#### Components
+
+**TinyScorer** (`scripts/rerank_recursive.py`)
+- 2-layer MLP neural network (~3MB per collection)
+- Scores query-document pairs based on learned patterns
+- Hot-reloads weights every 60 seconds from disk
+- Per-collection weights (each repo learns independently)
+
+**Event Logger** (`scripts/rerank_events.py`)
+- Logs every search to NDJSON files at `/tmp/rerank_events/`
+- Records: query, candidates, initial scores, timestamps
+- Hourly file rotation with configurable retention
+
+**Learning Worker** (`scripts/learning_reranker_worker.py`)
+- Background daemon that processes logged events
+- Uses ONNX cross-encoder as "teacher" model
+- Trains TinyScorer via knowledge distillation
+- Saves versioned weight checkpoints atomically
+
+#### Learning Flow
+
+1. **Event Capture**: Every search logs query + candidates to NDJSON
+2. **Teacher Scoring**: ONNX cross-encoder scores the candidates
+3. **Student Training**: TinyScorer learns to match teacher rankings
+4. **Weight Update**: New weights saved atomically with versioning
+5. **Hot Reload**: Serving path picks up new weights within 60s
+6. **Score Integration**: `learning_score` blends with other signals
+
+#### Configuration
+
+| Variable | Description | Default |
+|----------|-------------|---------|
+| `RERANKER_WEIGHTS_DIR` | Directory for weight files | `/tmp/rerank_weights` |
+| `RERANKER_WEIGHTS_RELOAD_INTERVAL` | Hot-reload check interval (seconds) | 60 |
+| `RERANKER_MAX_CHECKPOINTS` | Number of weight versions to keep | 5 |
+| `RERANKER_LR_DECAY_STEPS` | Steps between learning rate decay | 1000 |
+| `RERANKER_LR_DECAY_RATE` | Learning rate decay multiplier | 0.95 |
+| `RERANKER_MIN_LR` | Minimum learning rate | 0.0001 |
+| `RERANK_EVENTS_DIR` | Directory for event logs | `/tmp/rerank_events` |
+| `RERANK_EVENTS_RETENTION_DAYS` | Days to keep event files | 7 |
+| `RERANK_LEARNING_BATCH_SIZE` | Events per training batch | 32 |
+| `RERANK_LEARNING_POLL_INTERVAL` | Worker poll interval (seconds) | 30 |
+| `RERANK_LEARNING_RATE` | Initial learning rate | 0.001 |
+
+#### Observability
+
+Search results include learning metrics in the `why` field:
+```json
+{
+  "score": 3.2,
+  "why": ["lexical:1.0", "dense_rrf:0.05", "learning:3", "score:3.2"],
+  "components": {
+    "learning_score": 3.2,
+    "learning_iterations": 3
+  }
+}
+```
+
+Worker logs show training progress:
+```
+[codebase] Processed 5 events | v12 | lr=0.001 | avg_loss=1.8 | converged=False
+```
+
+#### Benefits
+
+- **Zero Manual Training**: Learns automatically from usage
+- **Per-Collection Specialization**: Each codebase gets tuned rankings
+- **Fast Inference**: TinyScorer adds <1ms to search latency
+- **Continuous Improvement**: Rankings improve over time
+- **Offline Capable**: Teacher runs locally, no external API calls
+
 #### MCP Router (`scripts/mcp_router.py`)
 - **Intent Classification**: Determines which MCP tool to call based on query
 - **Tool Orchestration**: Routes to search, answer, memory, or index tools
diff --git a/docs/CONFIGURATION.md b/docs/CONFIGURATION.md
index c2e93b7e..b7f90ca4 100644
--- a/docs/CONFIGURATION.md
+++ b/docs/CONFIGURATION.md
@@ -13,6 +13,7 @@ Complete environment variable reference for Context Engine.
 - [Query Optimization](#query-optimization)
 - [Watcher Settings](#watcher-settings)
 - [Reranker](#reranker)
+- [Learning Reranker](#learning-reranker)
 - [Decoder (llama.cpp / GLM / MiniMax)](#decoder-llamacpp--glm--minimax)
 - [ReFRAG](#refrag)
 - [Ports](#ports)
@@ -128,6 +129,61 @@ Dynamic HNSW_EF tuning and intelligent query routing for 2x faster simple querie
 | RERANKER_TOKENIZER_PATH | Tokenizer path for reranker | unset |
 | RERANKER_ENABLED | Enable reranker by default | 1 (enabled) |
 
+## Learning Reranker
+
+The learning reranker trains a lightweight neural network (TinyScorer) to improve search rankings over time. See [Architecture](ARCHITECTURE.md#5-learning-reranker-system) for details.
+
+**This feature is optional and enabled by default.** To disable:
+
+```bash
+# Disable learning scorer in search results
+RERANK_LEARNING=0
+
+# Disable event logging (no training data collected)
+RERANK_EVENTS_ENABLED=0
+
+# Or simply don't run the learning_worker container
+```
+
+### Enable/Disable
+
+| Name | Description | Default |
+|------|-------------|---------|
+| RERANK_LEARNING | Enable learning scorer in search results | 1 (enabled) |
+| RERANK_EVENTS_ENABLED | Enable event logging for training | 1 (enabled) |
+| RERANK_EVENTS_SAMPLE_RATE | Fraction of events to log (0.0-1.0) | 0.33 |
+
+### Weight Management
+
+| Name | Description | Default |
+|------|-------------|---------|
+| RERANKER_WEIGHTS_DIR | Directory for learned weight files | /tmp/rerank_weights |
+| RERANKER_WEIGHTS_RELOAD_INTERVAL | How often to check for new weights (seconds) | 60 |
+| RERANKER_MAX_CHECKPOINTS | Number of weight versions to retain | 5 |
+
+### Learning Rate
+
+| Name | Description | Default |
+|------|-------------|---------|
+| RERANKER_LR_DECAY_STEPS | Updates between learning rate decay | 1000 |
+| RERANKER_LR_DECAY_RATE | Decay multiplier (e.g., 0.95 = 5% reduction) | 0.95 |
+| RERANKER_MIN_LR | Minimum learning rate floor | 0.0001 |
+
+### Event Logging
+
+| Name | Description | Default |
+|------|-------------|---------|
+| RERANK_EVENTS_DIR | Directory for search event logs | /tmp/rerank_events |
+| RERANK_EVENTS_RETENTION_DAYS | Days to keep event files before cleanup | 7 |
+
+### Learning Worker
+
+| Name | Description | Default |
+|------|-------------|---------|
+| RERANK_LEARNING_BATCH_SIZE | Number of events per training batch | 32 |
+| RERANK_LEARNING_POLL_INTERVAL | Seconds between checking for new events | 30 |
+| RERANK_LEARNING_RATE | Initial learning rate for TinyScorer | 0.001 |
+
 ## Decoder (llama.cpp / GLM / MiniMax)
 
 | Name | Description | Default |
diff --git a/scripts/rerank_recursive.py b/scripts/rerank_recursive.py
index 8a886e55..84915922 100644
--- a/scripts/rerank_recursive.py
+++ b/scripts/rerank_recursive.py
@@ -1012,7 +1012,11 @@ def rerank_with_learning(
                     teacher_scores = None
 
         try:
-            from scripts.rerank_events import log_training_event
+            # Try both import paths for Docker (/app/scripts) and local (scripts/)
+            try:
+                from rerank_events import log_training_event
+            except ImportError:
+                from scripts.rerank_events import log_training_event
             log_training_event(
                 query=query,
                 candidates=candidates,