What it does

The parent orchestrator calls MemoryManager.prefetch_all() → kb_search(top_k=30) to Lore (CT 121) synchronously at the start of every run_conversation() (agent/conversation_loop.py ~654), before the first LLM call. Measured ~4.3s on every request.

The parent never uses that result for routing: SOUL.md explicitly delegates ALL recall to the memory child profile. The prefetch is therefore pure latency the orchestrator pays and discards.

This PR passes the existing skip_memory=True constructor flag at every user-facing PARENT/orchestrator construction point. When set, agent_init.py leaves _memory_manager=None and the prefetch is never reached. No change to conversation_loop.py or memory_manager.py — the prefetch logic is untouched; we simply don't build a memory manager on the parent.

Why it's safe

• SOUL delegates recall to the memory child — the parent never consults prefetch output for routing decisions, so dropping it changes no routing behavior.
• Children are unaffected. Delegated children construct their own _memory_manager independently (the generic delegate_task path in tools/delegate_tool.py and profile-based children decide their own skip_memory). Setting the flag on the parent does not flow into the child constructor — recall still works when the parent delegates to a memory-capable child.
• Explicit writes survive. memory(action="add") and kb_add via the MCP knowledge server are independent of the parent's _memory_manager prefetch path and continue to work.
• The cached gateway agent retains _memory_manager=None across turns (the cache stores the constructed instance; _init_cached_agent_for_turn never rebuilds the manager), so the saving applies to every turn in a session, not just the first.

Tradeoff surfaced for review — passive memory writes

The parent's _memory_manager is also what performs the post-turn passive transcript write:

conversation_loop.py (~4261) → AIAgent._sync_external_memory_for_turn (run_agent.py ~2104) → MemoryManager.sync_all → provider.sync_turn, plus queue_prefetch_all to warm the next turn.

With skip_memory=True, these passive writes no longer fire for PARENT-handled turns. Concretely, what stops accumulating in Lore from the parent path: the automatic (user, assistant) turn-pair transcript capture and the next-turn prefetch warming.

What is not lost:

• Explicit memory(action="add") writes and kb_add via the MCP knowledge server.
• Passive turn sync performed by delegated children (each child that builds its own memory manager still syncs its own turns).
• Curator / background-review already runs with skip_memory=True (and re-binds built-in MEMORY.md/USER.md from the parent), so it is unchanged.

Net: passive parent-transcript accumulation into Lore stops; child-driven memory and all explicit writes continue. This is the one real behavioral change — flagged for a human to accept before deploy.

Parent construction points changed

Construction points intentionally not changed:

• tools/delegate_tool.py _build_child_agent — already skip_memory=True (children); left as-is so child memory scoping is unchanged.
• gateway/platforms/feishu_comment.py, gateway _handle_compress_command, the gateway hygiene/compaction agent, background_review/curator — already skip_memory=True.

Flagged for reviewer

• ACP adapter (acp_adapter/session.py ~624) is another user-facing top-level parent that currently builds a memory manager and pays the same ~4.3s prefetch. It was outside the stated scope (api_server / telegram / TUI) so it is left unchanged here — call out if you want it included.

Tests

• tests/gateway/test_api_server_toolset.py — assert _create_agent passes skip_memory=True.
• tests/gateway/test_orchestrator_skip_memory.py (new) — assert the TUI parent helpers pass skip_memory=True; assert a real AIAgent(skip_memory=True) yields _memory_manager is None; assert the delegated-child construction carries its own skip_memory kwarg (parent value does not leak into the child).

Targeted run: 17 passed. Full tests/gateway/ suite: 5862 passed with 9 pre-existing, unrelated failures (telegram-markdown-escaping + wecom-callback test-isolation flakes) that reproduce identically on main without this change.

🤖 Generated with Claude Code