diff --git a/serve.py b/serve.py
new file mode 100644
index 0000000..2e58c3b
--- /dev/null
+++ b/serve.py
@@ -0,0 +1,97 @@
+#!/usr/bin/env python3
+"""Minimal CRI server. Exposes teach/trigger/generate endpoints.
+
+Usage:
+    pip install fastapi uvicorn
+    python serve.py --model Qwen/Qwen2.5-0.5B --port 8811
+
+API:
+    POST /teach    {"prompt": "...", "answer": "..."}
+    POST /trigger  {"query": "...", "max_tokens": 10}
+    POST /save     {"path": "bank.json"}
+    POST /load     {"path": "bank.json"}
+    GET  /stats
+"""
+
+import argparse
+import sys
+sys.path.insert(0, "python")
+
+from fastapi import FastAPI
+from pydantic import BaseModel
+import uvicorn
+
+from epimem import TransformersBackbone, EpisodicMemory, teach_fact, recall_fact
+
+app = FastAPI(title="CRI Server")
+backbone = None
+memory = None
+
+
+class TeachRequest(BaseModel):
+    prompt: str
+    answer: str
+
+class TriggerRequest(BaseModel):
+    query: str
+    max_tokens: int = 10
+    threshold: float = 0.3
+
+class PathRequest(BaseModel):
+    path: str
+
+
+@app.post("/teach")
+def teach(req: TeachRequest):
+    teach_fact(backbone, memory, req.prompt, req.answer)
+    return {"status": "conditioned", "total_reflexes": len(memory.episodes)}
+
+
+@app.post("/trigger")
+def trigger(req: TriggerRequest):
+    text, sim, episode = recall_fact(backbone, memory, req.query, req.max_tokens)
+    return {
+        "text": text.strip(),
+        "similarity": round(sim, 4),
+        "triggered": episode is not None,
+        "matched_prompt": episode["prompt"] if episode else None,
+    }
+
+
+@app.post("/save")
+def save(req: PathRequest):
+    memory.save(req.path)
+    return {"status": "saved", "path": req.path}
+
+
+@app.post("/load")
+def load(req: PathRequest):
+    memory.load(req.path)
+    return {"status": "loaded", "episodes": len(memory.episodes)}
+
+
+@app.get("/stats")
+def stats():
+    return {
+        "model": backbone.model_name,
+        "hidden_dim": backbone.hidden_dim,
+        "vocab_size": backbone.vocab_size,
+        "reflexes": len(memory.episodes),
+    }
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--model", default="Qwen/Qwen2.5-0.5B")
+    parser.add_argument("--port", type=int, default=8811)
+    parser.add_argument("--bank", type=str, default=None, help="Load reflex bank on startup")
+    args = parser.parse_args()
+
+    print(f"Starting CRI server on port {args.port}...")
+    backbone = TransformersBackbone(args.model)
+    memory = EpisodicMemory()
+
+    if args.bank:
+        memory.load(args.bank)
+
+    uvicorn.run(app, host="0.0.0.0", port=args.port)