Why does Manus forget my research notes?

Manus forgets your research notes because every task runs inside a fresh sandbox that resets on completion. Browsing history, scraped pages, and intermediate notes live as files inside that sandbox and disappear with it. Only the final artifact survives. The fix is a persistent research store that Manus reads from at the start of every new task.

Manus is engineered for long single tasks, not long projects. Three architectural choices cause the forgetting:

1. Research lives in sandbox files, not in memory. As the Manus team has explained, raw search results are saved to the agent's file system rather than held in the active prompt, which keeps the context lean enough to chain 50+ tool calls. Those files exist only inside that sandbox. When the task ends, the file system is gone with it.

2. The next task starts from your prompt, nothing more. The Manus engineers describe this directly: the 50th task you give the agent has the same starting context as the 1st. There is no built-in carry-over of what was already researched, ruled out, or sourced.

3. Older observations get summarized away inside one task too. Even within a single long task, Manus aggressively compresses older tool outputs to keep attention on the current step. Specific quotes and source URLs are the first things to fall out of the active window.

The result: Manus does excellent research inside a task, then loses the institutional memory of that research the moment the task closes.

Every new research task in the same project costs you the same kind of rework, and it stacks up fast:

• Sources get re-scraped. The 80 pages Manus already browsed last week get visited again, costing tool calls, time, and rate-limit headroom on the same sites.
• Vetting work resets. The judgement calls — which sources you ranked as credible, which you flagged as biased, which you dismissed — vanish, and Manus treats every source as new on the next task.
• Cross-source synthesis evaporates. The reconciled view across five conflicting reports lives only in the final artifact's prose, not as structured memory the agent can re-reason over.

The fix is not "attach yesterday's report as a PDF". It is to store the research as memory the agent can query, not as a frozen output.

Manus has a few mechanisms that partly help with research persistence.

Sandbox file system. During a single task, Manus uses files as durable scratch memory, and that is a real strength. It is also single-task only. The directory does not survive.

Final deliverables (PDF, doc, code). You can save the artifact Manus returns and re-feed it as an attachment to the next task. This preserves the conclusions but not the reasoning trace — which sources got cited, which got rejected, which alternative angles got explored before the final cut.

Manual context paste. The common workaround is to paste a long "what we already know" brief at the start of every Manus task. It works to a point. It also burns prompt budget that the agent could be spending on the next stretch of research, and the brief has to be maintained by hand as the project evolves.

For one-shot research, the natives are enough. For a research program that runs across many sessions, they are not.

The root issue is that Manus does not have a concept of a research project, only of a research task. There is no place inside the agent to store "the standing bibliography for this investigation" or "the list of dead-end hypotheses". Every task starts blind.

It gets worse when research moves between tools. You may collect sources with Manus, draft the report with Claude, fact-check with Perplexity, and polish with ChatGPT. Each tool has its own memory, none of them sees the others, and the canonical research record fragments across all of them.

MemoryLake is a cross-model memory layer that sits outside the agent sandbox. Your research lives in a Project, and Manus reads from that project at the start of every task instead of starting blank.

• A standing research bibliography per project. Vetted source URLs, key quotes, decisions about credibility, and ruled-out hypotheses are stored against the MemoryLake Project. Manus pulls the relevant slice into every new task automatically.
• 10,000× more context than raw prompting. MemoryLake's retrieval engine can hold billions of tokens of research per project and feed Manus only the relevant passages per task — far more than fits in any agent prompt, and without the per-task re-attachment dance.
• Portable across the rest of the research stack. The same project memory is readable by Perplexity, Claude, ChatGPT, Gemini, Grok, and any tool that speaks REST or MCP. When you switch to Claude to draft or Perplexity to fact-check, the bibliography is already there.

MemoryLake scored 94.03% on the LoCoMo long-context benchmark, the top result published as of 2026, with millisecond retrieval and AES-256 end-to-end encryption.

1. Create a project and load your research. Sign in to MemoryLake, open Project Management, click Create Project, and name it something like "Manus — competitor research". Upload prior reports, source lists, scraped pages, and notes through the Document Drive — PDF, Word, Excel, PowerPoint, Markdown, and images are all supported. Add structured entries for "credible sources" and "ruled-out hypotheses" in the Memories tab.
2. Generate an MCP Server endpoint. Open the MCP Servers tab inside your project, click Add MCP Server, name it "Manus research integration", and click Generate. MemoryLake returns an API key ID, secret, and endpoint URL. Copy the secret immediately — it is shown only once.
3. Connect Manus. Add MemoryLake as an MCP-compatible memory provider in the agent's tool/server config so Manus can query it during browsing tasks, or use the REST API with your Bearer token to fetch the project bibliography at task start and inject it into the opening prompt.