diff --git a/AGENTS.md b/AGENTS.md
index 9fe58fba..4546a7a3 100644
--- a/AGENTS.md
+++ b/AGENTS.md
@@ -13,22 +13,22 @@ fully-local Devstral model on qfox-1's RTX 4070. OpenClaw is decommissioned.
   - Run with `--insecure` (auth disabled) since the dashboard is on tailnet only.
   - Public access: `https://chat.researchstack.info` via Caddy on racknerd
     (`reverse_proxy http://100.85.244.73:9119`).
-- **Inference**: llama.cpp server on qfox-1 reachable at
-  `http://100.88.57.96:30516/v1` (k3s NodePort, ClusterIP unreachable from host)
-  - Model: `Devstral-Small-2505-UD-IQ2_M.gguf` (Unsloth, ~8.2 GB)
-  - Context: 131072 tokens (auto-detected as 16384, manually bumped in config)
-- **Hermes config**: `~/.hermes/config.yaml` uses the canonical schema:
+- **Local inference on qfox-1**: Ollama serving `hermes3:latest` built from
+  `/var/lib/gemma-models/Hermes-3-Llama-3.1-8B-Q4_K_M.gguf`.
+  - Endpoint: `http://127.0.0.1:11434/v1`
+  - Systemd unit: `ollama-hermes3.service`
+  - Note: `ollama-cuda` is installed but the packaged `llama-server` binary
+    currently falls back to CPU inference. GPU acceleration can be enabled later
+    by reinstalling Ollama from the official installer or building a CUDA-enabled
+    `llama-server`.
+- **Hermes config**: `~/.hermes/config.yaml` points at the local Ollama:
   ```yaml
   model:
-    default: "Devstral-Small-2505-UD-IQ2_M.gguf"
-    provider: "llamacpp"
-    base_url: "http://100.88.57.96:30516/v1"
+    default: "hermes3:latest"
+    provider: "openai"
+    base_url: "http://127.0.0.1:11434/v1"
     api_key: "sk-local"
-    context_length: 131072
   ```
-- **Ollama** (port 11434 on qfox-1) also has the same GGUF imported as
-  `devstral:latest` for ad-hoc `ollama run` use, but is **not** wired into
-  Hermes. Keep it as a debugging escape hatch.
 
 ### Common failure modes