refactor: update knowledge base policies and enhance memory loading logic in MemoryInjectionMiddleware

surfsense_backend/app/agents/new_chat/middleware/memory_injection.py

@@ -59,11 +59,12 @@ class MemoryInjectionMiddleware(AgentMiddleware):  # type: ignore[type-arg]
 
         async with shielded_async_session() as session:
             if self.user_id is not None:
-                user_memory = await self._load_user_memory(session)
+                user_memory, is_persisted = await self._load_user_memory(session)
                 if user_memory:
                     chars = len(user_memory)
+                    persisted = "true" if is_persisted else "false"
                     memory_blocks.append(
-                        f'<user_memory chars="{chars}" limit="{MEMORY_HARD_LIMIT}">\n'
+                        f'<user_memory chars="{chars}" limit="{MEMORY_HARD_LIMIT}" persisted="{persisted}">\n'
                         f"{user_memory}\n"
                         f"</user_memory>"
                     )
@@ -90,7 +91,16 @@ class MemoryInjectionMiddleware(AgentMiddleware):  # type: ignore[type-arg]
 
         return {"messages": new_messages}
 
-    async def _load_user_memory(self, session: AsyncSession) -> str | None:
+    async def _load_user_memory(
+        self, session: AsyncSession
+    ) -> tuple[str | None, bool]:
+        """Return (memory_content, is_persisted).
+
+        When the user has saved memory in the database, ``is_persisted`` is
+        ``True``.  When we fall back to a seed (first-name only), it is
+        ``False`` — the system prompt instructs the LLM to call
+        ``update_memory`` once to persist it.
+        """
         try:
             result = await session.execute(
                 select(User.memory_md, User.display_name).where(
@@ -99,25 +109,21 @@ class MemoryInjectionMiddleware(AgentMiddleware):  # type: ignore[type-arg]
             )
             row = result.one_or_none()
             if row is None:
-                return None
+                return None, True
 
             memory_md, display_name = row
 
             if memory_md:
-                return memory_md
+                return memory_md, True
 
-            # No saved memory yet — seed with the user's first name so the
-            # LLM knows who it's talking to from the very first turn.  The
-            # name is only injected, not persisted; the LLM will include it
-            # naturally when it first calls update_memory.
             if display_name:
                 first_name = display_name.split()[0]
-                return f"## About the user\n- Name: {first_name}"
+                return f"## About the user\n- Name: {first_name}", False
 
-            return None
+            return None, True
         except Exception:
             logger.exception("Failed to load user memory")
-            return None
+            return None, True
 
     async def _load_team_memory(self, session: AsyncSession) -> str | None:
         try:

surfsense_backend/app/agents/new_chat/system_prompt.py

@@ -257,12 +257,18 @@ _MEMORY_TOOL_INSTRUCTIONS: dict[str, dict[str, str]] = {
     `limit` attributes show your current usage and the maximum allowed size.
   - This is your curated long-term memory — the distilled essence of what you know about
     the user, not raw conversation logs.
-  - Call update_memory ONLY when the user shares genuinely important long-term information:
+  - **If <user_memory> has persisted="false"**, the memory has NOT been saved to the
+    database yet (it is a seed).  You MUST call update_memory during this conversation to
+    persist it, merging in any new information you learn from the user.
+  - Call update_memory when the user shares long-term information about themselves,
+    whether explicitly or casually as part of a request:
+    * Identity & background mentioned in passing: "I'm a student", "at my company we...",
+      "I'm working on a React app", "I'm a data scientist"
     * Explicit requests: "remember this", "keep in mind", "note that"
     * Preferences: "I prefer X", "I like Y", "always do Z"
     * Facts: name, role, company, expertise, current projects
     * Standing instructions: response format, communication style
-  - Do NOT call for trivial, temporary, or conversation-specific information.
+  - Skip truly ephemeral info (one-off questions, greetings, session logistics).
   - Args:
     - updated_memory: The FULL updated markdown document (not a diff).
       Merge new facts with existing ones, update contradictions, remove outdated entries.
@@ -281,12 +287,13 @@ _MEMORY_TOOL_INSTRUCTIONS: dict[str, dict[str, str]] = {
   - Your current team memory is already in <team_memory> in your context.  The `chars`
     and `limit` attributes show current usage and the maximum allowed size.
   - This is the team's curated long-term memory — decisions, conventions, key facts.
-  - Call update_memory ONLY when the team shares genuinely important long-term information:
-    * Team decisions: "let's remember we decided to use X"
+  - Call update_memory when the team shares long-term information, whether explicitly
+    or as part of a broader discussion:
+    * Team decisions: "let's use X", "we decided to go with Y"
     * Conventions: coding standards, processes, naming patterns
     * Key facts: where things are, how things work, team structure
     * Priorities: active projects, deadlines, blockers
-  - Do NOT call for trivial, temporary, or conversation-specific information.
+  - Skip truly ephemeral info (one-off questions, greetings, session logistics).
   - Args:
     - updated_memory: The FULL updated markdown document (not a diff).
       Merge new facts with existing ones, update contradictions, remove outdated entries.
@@ -305,12 +312,20 @@ _MEMORY_TOOL_INSTRUCTIONS: dict[str, dict[str, str]] = {
 _MEMORY_TOOL_EXAMPLES: dict[str, dict[str, str]] = {
     "update_memory": {
         "private": """
+- <user_memory persisted="false"> contains "## About the user\\n- Name: Alex"
+  User: "I'm a university student, explain astrophage to me"
+  - Memory is not yet persisted AND the user casually shared that they are a student.
+    You MUST call update_memory to persist the seed plus the new fact:
+    update_memory(updated_memory="## About the user\\n- Name: Alex\\n- University student\\n")
 - User: "Remember that I prefer TypeScript over JavaScript"
   - Timeless preference, no date needed.  You see the current <user_memory> and merge:
     update_memory(updated_memory="## About the user\\n- Senior developer\\n\\n## Preferences\\n- Prefers TypeScript over JavaScript\\n...")
 - User: "I actually moved to Google last month"
   - Fact that changes over time, include date:
     update_memory(updated_memory="## About the user\\n- Senior developer at Google (since 2026-03, previously Acme Corp)\\n...")
+- User: "I'm building a SaaS app with Next.js and Supabase"
+  - Implicit project info shared as context. Save it:
+    update_memory(updated_memory="## About the user\\n- Name: Alex\\n\\n## Current context\\n- Building a SaaS app with Next.js and Supabase (2026-04)\\n")
 """,
         "shared": """
 - User: "Let's remember that we decided to use GraphQL"