​

aiConnectedOS PRD Addendum

​

Feature: Conversation Split & Route

​

1. Overview

​

2. Feature Definition

​

2.1 What It Is

1. Receive a non-intrusive system notification when the AI detects a topic shift mid-conversation
2. View and modify which specific messages are proposed for extraction
3. Choose the destination folder or instance for the extracted content
4. Select whether to Move or Copy the selected messages
5. Have relevant memories from the extracted portion routed to the correct destination
6. Retain a visible, navigable link between the origin chat and the new destination chat

​

2.2 What It Is Not

• Replace or duplicate the existing Automatic Chat Cleanup function (Part 14), which handles whole-conversation relocation via cron-style background review
• Operate as a manual “move chat” tool (that already exists at the folder/instance level)
• Inject any notification or prompt into the conversation record itself
• Move an entire conversation — only a user-selected subset of messages within it

​

3. Trigger Conditions

​

3.1 Automatic Detection (AI-Initiated)

• The last N messages (configurable, suggested default: 4–6) are semantically divergent from the dominant topic of the conversation’s earlier content
• The new topic maps to an existing folder or instance in the user’s workspace with reasonable confidence (threshold: configurable, suggested default: 0.75 cosine similarity against folder-level metadata and Neurigraph node tags)
• The divergence has been sustained for at least 2 consecutive AI response cycles (to avoid false positives on brief tangents)

• Tangential questions that return to the main topic within 1–2 exchanges
• Meta-conversation (e.g., the user asking about the feature itself)
• Topics that have no identifiable destination folder or instance

​

3.2 Manual Initiation (User-Initiated)

• The ⋯ message-level action menu on any message (selecting “Split from here” sets that message as the proposed inflection point)
• The ChatNav sidebar, via a “Split conversation” action available on any checkpoint
• A context menu triggered by selecting multiple messages in the thread and choosing “Move selected” or “Copy selected”

​

4. The Overlay Notification

​

4.1 Design Principle

​

4.2 Notification Anatomy

• Icon indicating a topic shift has been detected
• Title: “Conversation has shifted topics”

• Brief one-line description of the detected shift, e.g.: “This conversation moved from [Personal / Wellness] into [aiConnected / Product Strategy].”
• Suggested destination: the folder or instance the system recommends, with its icon and name displayed

• Review & Route — opens the full Split Review panel (see Section 5)
• Dismiss — closes the notification with no action taken; the conversation continues normally

• Don't suggest this again for this chat — suppresses future shift notifications in the current conversation only
• Settings — deep link to notification sensitivity settings

​

4.3 Behavior

• The notification does not interrupt typing or message input
• It auto-dismisses after a configurable timeout (suggested default: 45 seconds) if the user takes no action, leaving no trace
• Dismissing it once does not permanently suppress it; a new and sufficiently different topic shift in the same chat will trigger it again
• The notification is not logged in the conversation history, the ChatNav index, or Neurigraph memory

​

5. The Split Review Panel

​

5.1 Message Selection View

• Deselect any pre-selected message to exclude it from the extraction
• Select any earlier message to include it in the extraction (useful if the topic shift actually began slightly before the AI detected it)
• Use “Select all after this message” as a shortcut on any message to extend selection forward from that point
• Use “Deselect all” to clear and start fresh

• At least one message must be selected to proceed
• The user can select non-contiguous messages (individual messages scattered throughout the thread)
• The selection does not need to be linear; users may cherry-pick specific exchanges

​

5.2 Destination Picker

• The system pre-fills its best suggested destination (the folder or instance it identified during detection)
• The user can change this to any accessible folder or instance via a searchable dropdown or folder tree
• The user can also choose “New folder…” or “New instance…” to create a destination on the spot, without leaving the flow
• The destination shows its current folder name, instance name, and any active persona assigned to it

​

5.3 Move vs. Copy Selection

• The selected messages no longer exist in the origin chat
• The origin chat record reads as if those messages never occurred there
• The user has a choice (see Section 7.1, Open Decision) about whether a subtle trace marker appears at the extraction point in the origin chat
• Memory nodes generated from the moved messages are transferred exclusively to the destination’s Neurigraph scope; they are removed from the origin scope

• Both conversations contain the selected messages
• The origin chat is unchanged
• Memory nodes generated from the copied messages exist in both the origin and destination Neurigraph scopes
• If the user wants to limit memory to only one scope, a prompt appears: “Should memories from these messages apply to both locations, or only the destination?” This gives the user explicit control rather than automatic duplication

​

5.4 Conversation Naming

​

5.5 Confirmation and Execution

​

6. Post-Operation State

​

6.1 The New Destination Chat

• Open with the selected messages as its initial content, in their original chronological order
• Display a “Branched from” banner at the top of the thread (consistent with the existing Linked Conversations pattern), e.g.: “Branched from [origin chat name] on [date]” with a View original link
• Have its own ChatNav index generated from the extracted messages
• Be scoped to the correct folder or instance’s persona, instruction layer, and memory context

​

6.2 The Origin Chat

• The selected messages are absent from the thread
• The conversation flows from the last unextracted message without visual disruption (default behavior)
• If the trace marker option is enabled (see Section 7.1), a subtle non-message system indicator appears at the extraction point reading: “N messages moved to [destination chat name]” with a jump link. This indicator is read-only, cannot be interacted with in the normal message sense, and does not affect context or memory retrieval.

• The origin chat is completely unchanged
• A subtle link annotation may optionally be added to the origin messages that were copied, indicating they also exist in the destination chat (mirroring the Linked Conversations link icon pattern)

​

6.3 Conversation Linking

type ConversationLink = {
  id: string;
  originChatId: string;
  destinationChatId: string;
  originMessageIds: string[];       // the specific messages that were extracted
  operationType: 'move' | 'copy';
  inflectionMessageId: string;      // the first message of the shift
  createdAt: string;
  createdBy: 'system' | 'user';     // whether the split was AI-initiated or manual
};

​

7. Memory Routing

​

7.1 Neurigraph Scope Assignment

​

7.2 Memory Integrity

1. Flags the node for review
2. Presents the user with a brief prompt: “This memory was built from messages in both locations. Which context should own it?” — with options: Origin, Destination, Both
3. Applies the user’s selection before closing the panel

​

8. ChatNav Integration

​

9. UX Notes

​

10. Acceptance Criteria

• System detects a topic shift with no false positives on tangential exchanges of 1–2 messages
• Shift notification appears as a UI overlay and is confirmed to write nothing to the conversation record, ChatNav index, or Neurigraph memory
• Notification auto-dismisses after configurable timeout without side effects
• Split Review panel pre-selects messages from the detected inflection point forward
• User can modify the selection (add, remove, non-contiguous selections) without constraints beyond the minimum of 1 message
• Destination picker surfaces the system’s best guess and allows full override
• “New folder” and “New instance” are accessible inline without leaving the panel
• Move operation removes selected messages from origin chat record
• Copy operation leaves origin chat fully intact
• Memory nodes are correctly re-scoped per the operation type and user choices
• Spanning memory nodes (across extracted and non-extracted messages) prompt the user for scope resolution
• Destination chat is created with the extracted messages in chronological order
• “Branched from” banner appears in the destination chat
• ConversationLink record is created and accessible from both chats’ header menus
• ChatNav in both chats reflects the post-split state correctly
• The full flow is completable on mobile
• The operation is fully reversible up to and excluding the Confirm action
• No part of the notification or panel UI becomes part of the conversation history

​

11. Dependencies

• ChatNav — checkpoint generation and update on split must be implemented and stable
• Linked Conversations — ConversationLink data model and header UI must be in place
• Folder System — destination picker relies on the folder/instance tree and move API
• Neurigraph Memory Architecture — memory re-scoping requires the partition model to support message-level provenance on nodes
• Automatic Chat Cleanup (Part 14) — shares topic classification logic; the detection engine for this feature should be built on the same classifier to avoid duplication
• Message-level action menu (⋯) — manual trigger requires the per-message action menu to support “Split from here” as a menu item

​

12. Implementation Notes

​

13. Open Decisions