From e26098237d3292b40ecd293c6444ff39066a0458 Mon Sep 17 00:00:00 2001
From: Anatoly Koptev <anatolykoptev@users.noreply.github.com>
Date: Fri, 6 Feb 2026 22:17:51 -0800
Subject: [PATCH] fix: activation memory config crashes get_default() with
 OpenAI backend

Two bugs in `get_default_config()` / `get_default_cube_config()`:

1. `get_default_config()` injects `act_mem` dict into MOSConfig when
   `enable_activation_memory=True`, but MOSConfig has no `act_mem` field
   and inherits `extra="forbid"` from BaseConfig. This causes a
   `ValidationError: Extra inputs are not permitted` for any user calling
   `get_default()` with activation memory enabled.

2. `get_default_cube_config()` hardcodes `extractor_llm` backend to
   `"openai"` for KV cache activation memory, but `KVCacheMemoryConfig`
   validator requires `huggingface`/`huggingface_singleton`/`vllm`
   (KV cache needs local model access for attention tensor extraction).
   This causes `ConfigurationError` even if bug #1 is fixed.

Fix: Remove `act_mem` from MOSConfig dict (the `enable_activation_memory`
bool flag is sufficient). In MemCube config, require explicit
`activation_memory_backend` kwarg instead of hardcoding `"openai"`.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
---
 src/memos/mem_os/utils/default_config.py | 56 +++++++++++++-----------
 1 file changed, 30 insertions(+), 26 deletions(-)

diff --git a/src/memos/mem_os/utils/default_config.py b/src/memos/mem_os/utils/default_config.py
index edb7875d4..9898cbe8c 100644
--- a/src/memos/mem_os/utils/default_config.py
+++ b/src/memos/mem_os/utils/default_config.py
@@ -3,12 +3,15 @@
 Provides simplified configuration generation for users.
 """
 
+import logging
 from typing import Literal
 
 from memos.configs.mem_cube import GeneralMemCubeConfig
 from memos.configs.mem_os import MOSConfig
 from memos.mem_cube.general import GeneralMemCube
 
+logger = logging.getLogger(__name__)
+
 
 def get_default_config(
     openai_api_key: str,
@@ -116,20 +119,9 @@ def get_default_config(
             },
         }
 
-    # Add activation memory if enabled
-    if config_dict.get("enable_activation_memory", False):
-        config_dict["act_mem"] = {
-            "backend": "kv_cache",
-            "config": {
-                "memory_filename": kwargs.get(
-                    "activation_memory_filename", "activation_memory.pickle"
-                ),
-                "extractor_llm": {
-                    "backend": "openai",
-                    "config": openai_config,
-                },
-            },
-        }
+    # Note: act_mem configuration belongs in MemCube config (get_default_cube_config),
+    # not in MOSConfig which doesn't have an act_mem field (extra="forbid").
+    # The enable_activation_memory flag above is sufficient for MOSConfig.
 
     return MOSConfig(**config_dict)
 
@@ -237,21 +229,33 @@ def get_default_cube_config(
             },
         }
 
-    # Configure activation memory if enabled
+    # Configure activation memory if enabled.
+    # KV cache activation memory requires a local HuggingFace/vLLM model (it
+    # extracts internal attention KV tensors via build_kv_cache), so it cannot
+    # work with remote API backends like OpenAI.
+    # Only create act_mem when activation_memory_backend is explicitly provided.
     act_mem_config = {}
     if kwargs.get("enable_activation_memory", False):
-        act_mem_config = {
-            "backend": "kv_cache",
-            "config": {
-                "memory_filename": kwargs.get(
-                    "activation_memory_filename", "activation_memory.pickle"
-                ),
-                "extractor_llm": {
-                    "backend": "openai",
-                    "config": openai_config,
+        extractor_backend = kwargs.get("activation_memory_backend")
+        if extractor_backend in ("huggingface", "huggingface_singleton", "vllm"):
+            act_mem_config = {
+                "backend": "kv_cache",
+                "config": {
+                    "memory_filename": kwargs.get(
+                        "activation_memory_filename", "activation_memory.pickle"
+                    ),
+                    "extractor_llm": {
+                        "backend": extractor_backend,
+                        "config": kwargs.get("activation_memory_llm_config", {}),
+                    },
                 },
-            },
-        }
+            }
+        else:
+            logger.info(
+                "Activation memory (kv_cache) requires a local model backend "
+                "(huggingface/vllm) via activation_memory_backend kwarg. "
+                "Skipping act_mem in MemCube config."
+            )
 
     # Create MemCube configuration
     cube_config_dict = {