Why does Manus forget my project history?

Manus forgets your project history because every task runs inside a fresh sandbox that resets when the task completes. The agent keeps a per-task todo.md and offloads work to files inside the sandbox, but none of it survives to the next run. Project history lives only in your final output artifacts, not in the agent. A persistent memory layer outside the sandbox fixes this.

Manus is built around a per-task sandbox. The Manus engineering team has been open about the trade-off: keeping state lean is what lets a single task chain 50+ tool calls without the context collapsing.

1. The sandbox resets at the end of every task. As the Manus team has written, the 50th task you give the agent starts with the same context as the 1st — your prompt and nothing more. The browsing history, the code it wrote, the search results it gathered, and the synthesis it produced all live inside the sandbox and disappear when the task completes.

2. In-context memory is deliberately recycled. To avoid attention collapse on long horizons, Manus rewrites its todo.md to keep the current objective near the end of the context. Older observations are summarized aggressively or dropped. That works inside one task. It guarantees forgetting across tasks.

3. Files are the agent's "memory", but only for that run. Manus stores raw search results, intermediate code, and large outputs as files inside the sandbox rather than keeping them in the live prompt. Smart for the current task, useless for the next one, because the file system goes with the sandbox.

The architectural choice that keeps Manus fast inside a single task is the same choice that makes it forget across them.

Each new task in a multi-week project costs you the same setup tax, and the loss compounds:

• Dead ends get re-walked. The three approaches Manus already tried and ruled out get tried again, burning tool calls and your patience.
• Research has to be redone. The 200 pages it browsed last week are gone. Same queries, same scraping, same summarization, billed to you again.
• Conventions reset. File naming, folder layout, framework choices, and the small decisions you settled on with Manus on day one all need to be re-stated in every new prompt.

The fix is not "write a longer prompt". It is to put project history somewhere outside the sandbox, where every new task can read it.

Manus ships a few mechanisms that partly help.

todo.md and in-task file memory. Excellent for the current task. The agent uses its own file system as scratch space and recites objectives to keep them in focus. None of this survives a sandbox reset.

Final output artifacts. Manus returns deliverables (documents, code, reports) you can save and re-feed into the next task as attachments. This works for stable outputs, but the agent's reasoning trace — why it picked option A over B, which sources it trusted, which paths it abandoned — is not in the artifact.

Manual prompt stuffing. Many users paste a long "project brief" at the top of every Manus task. This is the de facto workaround. It hits the same ceiling as every other long-prompt strategy: the more you paste, the less budget Manus has for the actual task, and you still have to maintain that brief by hand.

For one-off automations, the natives are fine. For multi-task projects that span weeks, they are not.

The deeper problem is that the agent has no notion of a project. It has tasks. Each task is independent. So even when you treat ten consecutive runs as a single piece of work, Manus does not — and there is no place inside Manus to put the shared context.

Worse, project work rarely stays inside Manus. You research in Manus, draft in ChatGPT, review in Claude, ship in Cursor. Each tool forgets in its own way, and the project context fragments across every one of them.

MemoryLake is a cross-model memory layer that lives outside any agent's sandbox. You create a project once, load it with your existing artifacts and decisions, and Manus pulls from that project at the start of every task.

• Per-project memory that outlives the sandbox. Briefs, prior outputs, research notes, ruled-out approaches, and naming conventions are stored against the MemoryLake Project. When Manus starts task 51, the first thing it loads is the project memory, not a blank prompt.
• 10,000× more context than raw prompting. MemoryLake's retrieval engine reads from billions of tokens of project memory and feeds Manus only what is relevant per task, so you stop paying tokens to paste the same brief every time and stop blowing out the sandbox's working context.
• Portable to the rest of your stack. The same project memory is readable by ChatGPT, Claude, Gemini, Grok, Cursor, and any tool that speaks REST or MCP. When the work moves out of Manus into review or shipping, the project history moves with it.

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 context. Sign in to MemoryLake, open Project Management, click Create Project, and name it something like "Manus — Q2 market research". Upload prior task outputs, briefs, decision logs, and reference docs through the Document Drive — PDF, Word, Excel, PowerPoint, Markdown, and images are all supported. Add naming conventions and ruled-out approaches as structured entries 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 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 call it at task initialization, or use the REST API with your Bearer token to fetch project memory programmatically at the start of every task and inject it into the opening prompt.