Ad Hoc Research

Missing chat context breaks research continuity in two different ways

A deep research task that starts without the originating chat context is likely to drift for two separate reasons: some products treat research as a partially separate artifact, and long multi-turn conversations themselves degrade model reliability.

What the sources say

• Direct evidence of separate-artifact behavior: an OpenAI community bug report says a completed deep research result can vanish from the visible thread after a follow-up. The opening report says this happened "3/3 times in the past 24 hours" on an Enterprise account and that "the original deep research response disappears" after a follow-up. URL: https://community.openai.com/t/deep-research-response-disappears-after-follow-up/1360916
• The same thread includes a plausible architecture clue from a quoted explanation: deep research results may be "ephemeral" and may "run in a separate research thread." This is not an official OpenAI product spec, but it matches the observed behavior that the report can remain accessible elsewhere while disappearing from the main view. URL: https://community.openai.com/t/deep-research-response-disappears-after-follow-up/1360916
• Anthropic’s memory design shows one mitigation pattern: Claude "will automatically summarize your conversations and create a synthesis of key insights across your chat history," that synthesis is "updated every 24 hours," and it provides context for "every new standalone conversation." Anthropic also explicitly limits the scope to chat history "not including chats in projects." URL: https://support.claude.com/en/articles/11817273-use-claude-s-chat-search-and-memory-to-build-on-previous-context
• Microsoft’s Copilot guidance treats context control as an explicit user action, not something the model can infer safely forever. It recommends: "Use threads to start a new conversation for a new task" and "Delete requests that are no longer relevant or that didn’t give you the desired result." URL: https://learn.microsoft.com/en-us/visualstudio/ide/copilot-chat-context-references?view=visualstudio
• OpenAI’s own Codex prompting guide points to the same pattern at API level. It says: "Compaction enables multi-hour reasoning without hitting context limits and longer continuous user conversations without needing to start new chat sessions." It also says compaction lets the model "retain key prior state with fewer conversation tokens." URL: https://developers.openai.com/cookbook/examples/gpt-5/codex_prompting_guide
• Independent technical analysis argues that raw context size is not the answer. Drew Breunig writes: "Just because there’s a 1 million token context window doesn’t mean you should fill it." He also gives a concrete overload example: one evaluation failed when a model saw "46 different tools" but succeeded when it only saw "19 tools." URL: https://www.dbreunig.com/2025/06/22/how-contexts-fail-and-how-to-fix-them.html
• The strongest quantitative result comes from the paper "LLMs Get Lost In Multi-Turn Conversation." Across "200,000+ simulated conversations," models showed an "average drop of 39% across six generation tasks" in multi-turn settings versus single-turn. Its summary conclusion is the clearest explanation for your situation: "when LLMs take a wrong turn in a conversation, they get lost and do not recover." URL: https://arxiv.org/abs/2505.06120

Synthesis

• If the research task was launched without the original chat, the system may literally lack the missing facts, constraints, and hypotheses. In that case the failure is not subtle: the research agent is solving a different problem.
• Even if the original chat was partially available, long-turn accumulation still causes degradation. The arXiv paper’s 39% average drop and Breunig’s tool-overload example both suggest that more history can make answers worse when the wrong details stay in frame.
• The vendor documentation converges on the same operational rule: context should be curated, compacted, or segmented. Anthropic uses periodic synthesis, Microsoft uses threads and scoped references, and OpenAI recommends compaction.

Practical takeaway

• For deep research, the best handoff is not the full transcript. It is a compact brief that carries the original objective, constraints, definitions, non-goals, and the specific open questions the research should answer.
• If the original conversation matters, preserve it as a distilled context packet before starting the research run. Include the exact research question, key assumptions, success criteria, and any must-use sources.
• If follow-up work changes topic, start a new thread instead of piling more turns onto the same history. The Microsoft guidance and the multi-turn reliability paper both support that move.

Bottom line

The evidence does not support the idea that a deep-tier research task can safely recover from missing chat context on its own. The dominant pattern across product docs, user reports, and published evaluation is that missing or bloated context changes the task itself. The fix is explicit context packaging: summarize the originating conversation into a compact, scoped brief before launching or continuing research.