refac

backend/open_webui/utils/memory.py

@@ -0,0 +1,314 @@
+from __future__ import annotations
+
+import asyncio
+import json
+import logging
+from typing import Any
+
+from fastapi import HTTPException
+
+from open_webui.models.config import Config
+from open_webui.models.memories import Memories
+from open_webui.utils.misc import add_or_update_system_message, get_content_from_message
+
+log = logging.getLogger(__name__)
+
+MEMORY_CONTEXT_OPEN = '<memory_context>'
+MEMORY_CONTEXT_CLOSE = '</memory_context>'
+
+
+def clean_memory_content(content: str | None) -> str:
+    value = (content or '').strip()
+    if not value:
+        raise HTTPException(status_code=400, detail='Memory content cannot be empty')
+    return value
+
+
+def validate_memory_operations(form_data) -> list[dict]:
+    if not form_data.operations:
+        raise HTTPException(status_code=400, detail='No memory operations provided')
+
+    operations = []
+    for operation in form_data.operations:
+        op = operation.model_dump()
+        action = op.get('action')
+
+        if action == 'add':
+            op['content'] = clean_memory_content(op.get('content'))
+            op['type'] = Memories.normalize_memory_type(op.get('type'))
+        elif action == 'replace':
+            if not op.get('id'):
+                raise HTTPException(status_code=400, detail='Memory id is required for replace')
+            op['content'] = clean_memory_content(op.get('content'))
+            if op.get('type') is not None:
+                op['type'] = Memories.normalize_memory_type(op.get('type'))
+        elif action == 'remove':
+            if not op.get('id'):
+                raise HTTPException(status_code=400, detail='Memory id is required for remove')
+        else:
+            raise HTTPException(status_code=400, detail=f'Unsupported memory operation: {action}')
+
+        operations.append(op)
+
+    return operations
+
+
+def model_allows_memory(model: dict | None) -> bool:
+    return (model or {}).get('info', {}).get('meta', {}).get('capabilities', {}).get('memory', True)
+
+
+async def add_memory_context(request, form_data: dict, user, model: dict | None = None):
+    if not model_allows_memory(model):
+        return form_data
+
+    user_messages = []
+    for message in reversed(form_data.get('messages', [])):
+        if message.get('role') != 'user':
+            continue
+
+        content = get_content_from_message(message)
+        if isinstance(content, str) and content.strip():
+            user_messages.append(content.strip())
+
+        if len(user_messages) >= 7:
+            break
+
+    query = '\n\n'.join(reversed(user_messages))[-4000:]
+    if not query:
+        return form_data
+
+    try:
+        from open_webui.routers.memories import QueryMemoryForm, query_memory
+
+        results = await query_memory(request, QueryMemoryForm(content=query, k=6), user)
+    except Exception as e:
+        log.debug(e)
+        return form_data
+
+    sections = {'user': [], 'context': []}
+    if results and hasattr(results, 'documents') and results.documents:
+        for doc_idx, doc in enumerate(results.documents[0]):
+            if not doc:
+                continue
+
+            metadata = {}
+            if results.metadatas and results.metadatas[0] and len(results.metadatas[0]) > doc_idx:
+                metadata = results.metadatas[0][doc_idx] or {}
+
+            sections[Memories.normalize_memory_type(metadata.get('type'))].append(str(doc))
+
+    parts = []
+    if sections['user']:
+        parts.append('User:\n' + '\n'.join(f'- {memory}' for memory in sections['user']))
+    if sections['context']:
+        parts.append('Context:\n' + '\n'.join(f'- {memory}' for memory in sections['context']))
+    if not parts:
+        return form_data
+
+    limit = await Config.get('memories.context_char_limit', 2000)
+    try:
+        limit = max(250, int(limit))
+    except Exception:
+        limit = 2000
+
+    messages = form_data['messages']
+    if messages and messages[0].get('role') == 'system':
+        content = messages[0].get('content', '')
+        if isinstance(content, str) and MEMORY_CONTEXT_OPEN in content:
+            start = content.find(MEMORY_CONTEXT_OPEN)
+            end = content.find(MEMORY_CONTEXT_CLOSE, start)
+            if end != -1:
+                messages[0]['content'] = (content[:start] + content[end + len(MEMORY_CONTEXT_CLOSE) :]).strip()
+
+    memory_context = f'{MEMORY_CONTEXT_OPEN}\n' + '\n\n'.join(parts)[:limit] + f'\n{MEMORY_CONTEXT_CLOSE}'
+    form_data['messages'] = add_or_update_system_message(memory_context, messages, append=True)
+    return form_data
+
+
+async def review_memory_after_turn(
+    *,
+    request,
+    user,
+    model: dict | None,
+    metadata: dict,
+    form_data: dict,
+    assistant_message: dict,
+    messages: list[dict],
+) -> None:
+    if not model_allows_memory(model):
+        return
+
+    features = metadata.get('features') or {}
+    if not features.get('memory'):
+        return
+
+    assistant_content = assistant_message.get('content', '')
+    if not isinstance(assistant_content, str) or not assistant_content.strip():
+        return
+
+    config = await Config.get_many(
+        'memories.background_review.enable',
+        'memories.review_interval_turns',
+    )
+    if not config.get('memories.background_review.enable'):
+        return
+
+    try:
+        interval = max(1, int(config.get('memories.review_interval_turns', 10)))
+    except Exception:
+        interval = 10
+
+    user_turns = len([message for message in messages if message.get('role') == 'user'])
+    if user_turns == 0 or user_turns % interval != 0:
+        return
+
+    task = asyncio.create_task(
+        _review_memory(
+            request=request,
+            user=user,
+            model=model,
+            metadata=metadata,
+            form_data=form_data,
+            assistant_message=assistant_message,
+            messages=messages,
+        )
+    )
+
+    def log_failure(done_task):
+        try:
+            done_task.result()
+        except Exception as e:
+            log.debug(f'Memory review failed: {e}')
+
+    task.add_done_callback(log_failure)
+
+
+async def _review_memory(
+    *,
+    request,
+    user,
+    model: dict | None,
+    metadata: dict,
+    form_data: dict,
+    assistant_message: dict,
+    messages: list[dict],
+) -> None:
+    existing_memories = await Memories.get_memories_by_user_id(user.id)
+    existing_lines = [
+        f'- id={memory.id} type={memory.type} content={memory.content}'
+        for memory in (existing_memories or [])[:80]
+    ]
+
+    assistant_content = assistant_message.get('content', '')
+    if not isinstance(assistant_content, str):
+        assistant_content = get_content_from_message(assistant_message)
+
+    transcript_lines = []
+    for message in messages[-16:]:
+        role = message.get('role', '')
+        content = message.get('content', '')
+        if not isinstance(content, str):
+            content = get_content_from_message(message)
+        content = content.strip()
+        if role not in {'user', 'assistant'} or not content:
+            continue
+        if len(content) > 1600:
+            content = f'{content[:1000]}\n...(truncated)...\n{content[-400:]}'
+        transcript_lines.append(f'{role}: {content}')
+
+    if assistant_content.strip():
+        assistant_final = assistant_content.strip()
+        if len(assistant_final) > 1600:
+            assistant_final = f'{assistant_final[:1000]}\n...(truncated)...\n{assistant_final[-400:]}'
+        transcript_lines.append(f'assistant_final: {assistant_final}')
+
+    model_id = model.get('id') if isinstance(model, dict) else form_data.get('model')
+    operations = await _generate_memory_operations(
+        request=request,
+        user=user,
+        model_id=model_id,
+        metadata=metadata,
+        existing_text='\n'.join(existing_lines) if existing_lines else '(none)',
+        transcript='\n\n'.join(transcript_lines),
+    )
+    if operations:
+        from open_webui.routers.memories import UpdateMemoriesForm, update_memories
+
+        await update_memories(request, UpdateMemoriesForm(operations=operations), user)
+
+
+async def _generate_memory_operations(
+    *,
+    request,
+    user,
+    model_id: str,
+    metadata: dict,
+    existing_text: str,
+    transcript: str,
+) -> list[dict[str, Any]]:
+    from open_webui.utils.chat import generate_chat_completion
+
+    review_prompt = f"""Review the completed conversation turn and decide whether long-term memory should change.
+
+Memory types:
+- user: durable facts, preferences, or instructions about the user.
+- context: other durable context that may help future chats for this user account.
+
+Rules:
+- Save only information likely to matter in future chats.
+- Do not save secrets, credentials, transient task steps, or unsupported guesses.
+- Prefer replace/remove over duplicate add when an existing memory should change.
+- Do not invent type, status, trait, score, importance, or stability schemas.
+- Return only JSON in this shape:
+  {{"operations":[
+    {{"action":"add","type":"user|context","content":"..."}},
+    {{"action":"replace","id":"...","type":"user|context","content":"..."}},
+    {{"action":"remove","id":"..."}}
+  ]}}
+- Use an empty operations array if nothing should be remembered.
+
+Existing memories:
+{existing_text}
+
+Conversation:
+{transcript}
+"""
+
+    response = await generate_chat_completion(
+        request,
+        form_data={
+            'model': model_id,
+            'messages': [
+                {
+                    'role': 'system',
+                    'content': "You are Open WebUI's private memory reviewer. Return only valid JSON.",
+                },
+                {'role': 'user', 'content': review_prompt},
+            ],
+            'stream': False,
+            'metadata': {
+                'task': 'memory_review',
+                'chat_id': metadata.get('chat_id'),
+                'message_id': metadata.get('message_id'),
+            },
+        },
+        user=user,
+    )
+
+    if not isinstance(response, dict) or not response.get('choices'):
+        return []
+
+    response_message = response.get('choices', [{}])[0].get('message', {})
+    content = response_message.get('content') or response_message.get('reasoning_content') or ''
+    start = content.find('{')
+    end = content.rfind('}')
+    if start == -1 or end == -1 or end < start:
+        return []
+
+    try:
+        parsed = json.loads(content[start : end + 1])
+    except Exception:
+        return []
+
+    operations = parsed.get('operations') if isinstance(parsed, dict) else None
+    return operations if isinstance(operations, list) else []

backend/open_webui/utils/middleware.py

@@ -53,7 +53,6 @@ from open_webui.routers.images import (
     image_edits,
     image_generations,
 )
-from open_webui.routers.memories import QueryMemoryForm, query_memory
 from open_webui.routers.pipelines import (
     process_pipeline_inlet_filter,
     process_pipeline_outlet_filter,
@@ -90,6 +89,7 @@ from open_webui.utils.filter import (
 )
 
 from open_webui.utils.mcp.client import MCPClient
+from open_webui.utils.memory import add_memory_context, review_memory_after_turn
 from open_webui.utils.misc import (
     add_or_update_system_message,
     add_or_update_user_message,
@@ -147,6 +147,7 @@ DEFAULT_REASONING_TAGS = [
     ('<|begin_of_thought|>', '<|end_of_thought|>'),
     ('◁think▷', '◁/think▷'),
 ]
+
 DEFAULT_SOLUTION_TAGS = [('<|begin_of_solution|>', '<|end_of_solution|>')]
 DEFAULT_CODE_INTERPRETER_TAGS = [('<code_interpreter>', '</code_interpreter>')]
 
@@ -1457,42 +1458,6 @@ async def chat_completion_tools_handler(
 
     return body, {'sources': sources}
 
-
-async def chat_memory_handler(request: Request, form_data: dict, extra_params: dict, user):
-    try:
-        results = await query_memory(
-            request,
-            QueryMemoryForm(
-                **{
-                    'content': get_last_user_message(form_data['messages']) or '',
-                    'k': 3,
-                }
-            ),
-            user,
-        )
-    except Exception as e:
-        log.debug(e)
-        results = None
-
-    user_context = ''
-    if results and hasattr(results, 'documents'):
-        if results.documents and len(results.documents) > 0:
-            for doc_idx, doc in enumerate(results.documents[0]):
-                created_at_date = 'Unknown Date'
-
-                if results.metadatas[0][doc_idx].get('created_at'):
-                    created_at_timestamp = results.metadatas[0][doc_idx]['created_at']
-                    created_at_date = time.strftime('%Y-%m-%d', time.localtime(created_at_timestamp))
-
-                user_context += f'{doc_idx + 1}. [{created_at_date}] {doc}\n'
-
-    form_data['messages'] = add_or_update_system_message(
-        f'User Context:\n{user_context}\n', form_data['messages'], append=True
-    )
-
-    return form_data
-
-
 async def chat_web_search_handler(request: Request, form_data: dict, extra_params: dict, user):
     event_emitter = extra_params['__event_emitter__']
     await event_emitter(
@@ -2649,8 +2614,7 @@ async def process_chat_payload(request, form_data, user, metadata, model):
                 )
 
         if 'memory' in features and features['memory']:
-            # Skip forced memory injection when native FC is enabled - model can use memory tools
-            form_data = await chat_memory_handler(request, form_data, user)
+            form_data = await add_memory_context(request, form_data, user, model)
 
         if 'web_search' in features and features['web_search']:
             # Skip forced RAG web search when native FC is enabled - model can use web_search tool
@@ -2787,6 +2751,7 @@ async def process_chat_payload(request, form_data, user, metadata, model):
         'tool_ids': tool_ids,
         'terminal_id': terminal_id,
         'files': files,
+        'features': features,
     }
     form_data['metadata'] = metadata
 

backend/open_webui/utils/tools.py

@@ -82,6 +82,7 @@ from open_webui.tools.builtin import (
     toggle_automation,
     update_automation,
     update_calendar_event,
+    update_memory,
     update_task,
     view_channel_message,
     view_channel_thread,
@@ -527,19 +528,21 @@ async def get_builtin_tools(
     if is_builtin_tool_enabled('chats'):
         builtin_functions.extend([search_chats, view_chat])
 
-    # Add memory tools if builtin category enabled AND enabled for this chat
+    # Add memory tools when memory is enabled and the model allows this builtin category.
     if (
         is_builtin_tool_enabled('memory')
-        and (features.get('memory') or get_model_capability('memory', False))
+        and features.get('memory')
+        and get_model_capability('memory')
         and await has_user_permission('memories')
     ):
         builtin_functions.extend(
             [
                 search_memories,
+                list_memories,
+                update_memory,
                 add_memory,
                 replace_memory_content,
                 delete_memory,
-                list_memories,
             ]
         )