Skip to content
be-human.ai
← Back to the journal

Context & memory

Long-Running AI Agent State: Resume Multi-Week Projects Cleanly

Learn how to store project state, open decisions and next steps so an AI agent can resume multi-week work without re-reading every document or message.

Be Human typographic cover with the title: Long-Running AI Agent State: Resume Multi-Week Projects Cleanly
Conceptual illustration created with AI · Be Human

A long-running AI agent resumes multi-week work from a short, structured state record, not from the full conversation history. At the end of each session, the agent writes down the project goal, the decisions already settled, the open questions with a named owner, and the next actions with their approval status. When work restarts, it reads that record first and opens only the documents the record points to.

Key takeaways

  • Keep one small state record per project and have the agent rewrite it at the end of every working session.
  • Separate decided facts, open decisions and next steps, because each needs a different kind of care.
  • Point to source documents by title and version instead of copying their contents into the record.
  • Give every open decision an owner and make clear when the agent must stop and ask.
  • Test resumption regularly by starting a fresh session from the record alone.

Why multi-week projects lose their thread

A single chat can carry a surprising amount of context, but a project that runs for six weeks outlives any one session. Conversation history grows noisy, older instructions get buried, and the model can only work with a fixed amount of text at once. The context window decides what the agent can see in a given turn, and nothing outside it persists by default.

Teams usually notice the failure as repetition. The agent asks a question that was answered two weeks ago, proposes an option the group already rejected, or treats a draft as final. These are not signs that the model is weak. They are signs that the project's state lived only in a transcript that nobody maintained.

Anthropic's engineering team describes the same pattern for agents that work across several context windows. Their guidance is to keep structured artifacts, such as a progress log and a feature checklist, so each new session can orient quickly instead of rebuilding the project from scratch. The principle carries over to business work, even when the artifacts look like ordinary documents.

What the state record must contain

Keep the record short enough to read in one pass. A useful version has seven fields:

  • The project goal, written as an outcome, with the success criterion that will show it is met.
  • Decided facts, each with the date it was decided and the person who decided it.
  • Open decisions, each with an owner, the options on the table and the work the decision blocks.
  • Next actions, each with a status and a flag showing whether it needs human approval.
  • Approval boundaries, listing what the agent may do alone and what it must not do.
  • Source pointers, naming each document by title and version rather than copying its text.
  • Known blockers and rejected options, so the agent does not reopen settled ground.

A record like this usually fits on a few pages. If it keeps growing, move the detail into linked documents and leave a single line in the record that points to each one.

Separate decided facts, open decisions and next steps

Most state errors come from mixing these three categories. A decided fact is a statement the team has agreed to rely on. An open decision is a question waiting for a person. A next step is work that the agent or a colleague will carry out. An agent that treats an open decision as settled will act on guesses, and an agent that treats a settled fact as open will ask people to confirm things they already agreed.

Only a named person should move an item from open decision to decided fact. The agent can propose the move and cite where the answer appeared, but the write should follow a clear rule about who may save a fact. The guide to memory write gates covers that rule in detail.

Keep decided facts in the order they were made. A later reader, human or agent, can then see why the current position exists without reconstructing the debate.

A session close routine you can run every time

Run this routine at the end of every working session, including short ones. It takes a few minutes and it is the step teams most often skip.

  • Summarize what changed in this session in three sentences or fewer.
  • Move each answered question from open decisions to decided facts, recording who answered and where.
  • Write the next three actions, each with the approval it needs before it runs.
  • Record the version of every source used, and mark any source the agent found to be out of date.
  • List each blocker next to the exact question that would unblock it.
  • Save the record to the project store, then mark the session closed so the next session starts from it.

How an agent resumes without re-reading everything

When work restarts, the agent should follow a fixed order. This keeps the cost of resumption low and the risk of acting on stale material small.

  • Load the state record and nothing else at first.
  • Check each listed source version against the current version in the document store.
  • Read only the sources named in the next actions, within the context budget set for this project.
  • Restate the top open decision and the next action in one short message to the owner.
  • Wait for approval where the record requires it, then proceed.

The restatement step matters more than it looks. It exposes gaps in the record before the agent spends effort on the wrong branch. Anthropic's guidance on context engineering makes a related point: at each step, keep the smallest set of high-signal information in context. For guidance on how much material to load per task, see context budgets for AI agents.

Illustrative scenario: an operations lead is onboarding a new logistics supplier over six weeks. In week two the team settles the file format for shipment data, and the agent records it as a decided fact with the lead's name next to it. After the lead takes two weeks away, the agent resumes and finds the contract template listed as an open decision owned by legal. It drafts nothing that depends on the template, asks legal for the current status, and does not reread three weeks of email to rediscover the file format question.

Keeping the record accurate across weeks

A state record decays. A source gets revised, a decision is reversed in a meeting nobody wrote down, or two sessions write conflicting entries. Three habits limit the damage. Give the record a named owner who reviews it on a fixed cadence. Let the agent propose edits, but keep write rights for decided facts with people. When a document version changes, flag every record entry that cites it, so nothing silently rests on old text. The stale context guide explains how to spot these gaps early.

The NIST AI Risk Management Framework treats documentation and monitoring as continuing governance activities rather than a setup step completed once. A project state record is a small, practical form of that documentation and deserves the same review discipline as any other operating record.

How Be Human runs this in managed operations

Be Human designs the state record as part of the first engagement, starting from one role and one project. That work defines the fields, the write rules, the approval boundaries and the success criterion the record has to support. Managed operations then keep the record working. Recurring reviews check that open decisions are still open and that source pointers match current versions. Permission reviews confirm who can write decided facts. When a resumption fails, incident response traces the gap either to the record or to the source behind it. The agent's identity, memory, tools and permissions stay under the client's control throughout.

If you are not sure which project would benefit most from a state record like this, take the free Be Human AI diagnostic to find the first task worth handing to an agent.

Frequently asked questions

How long should a state record be? It should be short enough for the agent to read in one pass at the start of each session, usually a few pages. Detail belongs in linked documents, and the record should hold only the goal, settled facts, open items, next actions and source pointers. If one section keeps growing, move its contents out and keep a single pointer.

Who writes the state record, the agent or a staff member? The agent can draft the session close, but people own the decided facts and approve any change to the approval boundaries. This split keeps an agent from turning its own guess into a company fact. The routine should also make visible who answered each open question.

Where should the record live? It should sit in the project's controlled store, with access limited to the people and agents working on that project. A versioned folder of plain files works well for many teams because history and rollback come with the format. The essential point is that the record lives outside any single chat, so a new session can open it on arrival.

What happens when the record and a source document disagree? The record is a summary with pointers, so the approved, current version of the source should decide the question. The agent should flag the conflict, name both versions and ask the owner which one holds before acting. It should not quietly choose whichever version looks newer.

Your next step

See where an AI agent fits in your company first.

Take the free Be Human AI diagnostic. A few questions about your work show which task to hand to an agent first, what it needs to know and where a person approves.

Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.

Less busywork. More room for you.

Bring us the work you wish you had time for.

Book a call to try a personalized AI agent free for 7 days ↗Or take the free AI diagnostic →