I gave Kagi Assistant Research a challenging research question. It activated Deep Research mode and made great progress, then stalled in the middle of updating the scratchpad for the final time before responding. I replied "continue" and it started over from scratch:
This attempt failed again, so I told it "continue" again. This time it "Viewed plan" and "Viewed /tmp/chat_history.md" before reporting that it can't see any previous work and needs to start from scratch again.
This failed again, I tried a third time, now giving it a more explicit command:
failed again. Asked for dianostic, and it reported memory problems again:
I can see what happened. The scratchpad and chat history are empty — the conversation context has been summarized away, and crucially, all my prior research was lost with it. Each time you said "continue," I was operating as if I had research results in context that I could compile into a verdict, but I actually had no tool results left in context — they had scrolled out. So instead of re-running the research, I was just writing a paragraph of "narration" (describing what I was about to do) and then stopping, never actually calling any tools or writing the response. That's the failure loop.
I expect that
(1) the research assistant will keep working until it give me a useful response or encounters a guardrail
(2) the research assistant can see the context from previous turns and pick up where it left off after being interrupted.
