Restore Copilot-managed shared memory in V4 Agent Chat

Is your feature request related to a problem? Please describe.

Copilot already has a vault-local memory system: Recent Conversations.md, Saved Memories.md, and the settings enableRecentConversations, enableSavedMemory, and maxRecentConversations.

In V4, Agent Chat is the primary desktop experience, but Codex, Claude, and OpenCode run through the separate ACP path. That path deliberately bypasses Copilot's existing LangChain memory/envelope stack, so Agent Chat does not appear to read or update Copilot's memory files. The memory settings can remain enabled without indicating that they apply only to Quick Chat.

This leaves no Copilot-managed memory shared across Agent backends. The available substitutes are different things:

• AGENTS.md stores instructions, not learned memory.
• Skills provide capabilities, not memory.
• Projects hold manually selected context and are project-scoped.
• Full chat transcripts are history and contain too much noise to serve as curated memory.
• Backend-specific memory is not portable when switching between Codex, Claude, and OpenCode.

The result is surprising: moving from Quick Chat to the new primary Agent Chat silently stops the Copilot-owned memory layer from functioning.

Describe the solution you'd like

Provide one backend-neutral, vault-local Copilot memory store that works across Codex, Claude, and OpenCode Agent Chat sessions.

Ideally:

• New Agent Chat sessions receive the enabled Saved Memories plus a bounded, relevant summary of recent conversations.
• Completed Agent Chat conversations can update the same rolling memory using the existing retention setting.
• The same Copilot memory follows the user when they switch Agent backends.
• Existing memory toggles remain the source of user control.
• Memory stays inspectable and editable as Markdown in the vault.
• Curated summaries are used rather than injecting full chat transcripts.

If ACP cannot support this safely, the minimum fix would be to label the current memory controls Quick Chat only and clearly document that Agent Chat has no Copilot-managed cross-session memory.

Describe alternatives you've considered

• A memory.md note attached to every Project: manual, project-scoped, and easy to let drift.
• Putting facts in AGENTS.md: mixes personal/project memory with behavioral instructions.
• Depending on each backend's native memory: produces inconsistent context when switching agents.
• Adding saved chats to Project context: useful continuity, but still manual and transcript-heavy. Issue #3108 proposes this workflow but explicitly does not add chats automatically.

Additional context

The ACP design document says Agent mode should stay isolated from the existing LangChain model/tool/memory stack and should not inject Copilot's L4 conversation strip: https://github.com/logancyang/obsidian-copilot/blob/master/designdocs/todo/ACP_DESIGN.md

The older memory design says “Agent Mode” has full access to Recent Conversations and Saved Memories, but that describes the previous autonomous-agent implementation, not V4 ACP Agent Chat: https://github.com/logancyang/obsidian-copilot/blob/master/src/memory/memory-design.md

Current Agent Chat documentation recommends Projects or AGENTS.md for reusable context, but neither is a replacement for automatic shared memory: https://github.com/logancyang/obsidian-copilot/blob/master/docs/agent-mode-and-tools.md#context-and-history

Observed with Copilot 4.0.9.