Agent Memory Handoff: Pass the Work to the Next Assistant

An agent memory handoff transfers the current goal, decisions, completed work, supporting evidence, and next action to the assistant taking over. Save those facts in a shared record, then have the receiving assistant read and verify them before continuing. The aim is to preserve the state of the work, including what remains uncertain.

Tools like fenbs keep that record outside any one AI app, so a new conversation can start from the same project progress. This guide covers the transfer itself; the AI handoffs hub covers related ways to carry work between tools and sessions.

How do you prepare an agent memory handoff?

1. Pause at a clear boundary. Identify which task is changing hands and whether the current assistant has stopped making changes.
2. Write the goal and the current state. Separate work completed from work attempted, planned, or waiting for approval.
3. Save the decisions that constrain the next action, including who made them and where the original agreement can be checked.
4. Attach evidence: the relevant document, task, file, test result, or source link. State what each reference proves.
5. Name the next action and ask the receiving assistant to explain its starting point before it acts.

Do not assume that a transfer between assistants includes every earlier message. Microsoft's AutoGen handoff documentation↗ (opens in a new tab) describes one arrangement where participating agents share message context. A new conversation in another app may have a different starting point. Check what the recipient can actually read.

What information belongs in the handoff?

Include information that changes what the next assistant should do. Keep the main handoff short enough to review, and link to detailed material instead of copying it all. The introduction to agent memory explains the broader idea of retaining information beyond one exchange.

For example, a design handoff might say that a revised signup page is drafted and reviewed for wording, while keyboard navigation remains unchecked. The next assistant should inspect that behavior before declaring the page ready. Saying only that the page is finished would erase the outstanding work.

Google's Agent Development Kit documentation↗ (opens in a new tab) distinguishes a conversation session from information retained across sessions. That distinction helps explain why a useful handoff needs a lasting home: being present in the old conversation does not establish that information is available in the new one.

How do you preserve decisions and evidence accurately?

Record a person's choice separately from an assistant's suggestion. Include its scope: a rule for one customer proposal should not quietly become a rule for every proposal. If nobody approved a recommendation, label it as an open question and name the person who can resolve it.

Use precise evidence language. “The draft was checked against the approved outline” tells the next assistant something useful. “Verified” alone does not say whether someone checked the facts, formatting, links, or approval. Record failed checks and checks you could not perform with the same care as successful ones.

Give references enough detail to find the right version. For a document, use its title and relevant section. For code, include the file and the change being discussed. If a source may change, note which claim you checked and when, without presenting an old result as a fresh check.

In fenbs, tasks hold the work while Decisions and rules preserve what people chose; the handoff workflow connects those records to Where we left off.

What should the receiving assistant check first?

Ask it to restate the goal, the most important constraint, and the first action it intends to take. This small readback exposes missing context before the assistant changes anything. If its answer disagrees with the current record, resolve the mismatch before proceeding.

Next, have it open the references needed for that action. A summary can direct the work, but it should not replace an approval or a source document when the exact wording matters. The difference between agent memory and context matters here: saved information helps only when the assistant receives and uses the relevant part.

Confirm access as well. A link to a private file does not give the next person or assistant permission to open it. Resolve access through the source system; avoid copying confidential material into a wider record just to make the handoff convenient.

Why do otherwise detailed handoffs fail?

A long transcript can bury the latest state beneath abandoned ideas. Anthropic's context engineering guidance↗ (opens in a new tab) recommends selecting relevant information rather than treating all available material as equally useful. For a handoff, preserve the reason behind the current choice and keep older discussion available by reference.

• Two assistants continue changing the same task. Name who is taking over and stop overlapping work before transferring it.
• The summary sounds more certain than the evidence. Keep assumptions, untested work, and unresolved questions explicit.
• The handoff is saved in a private conversation. Put the reviewed record somewhere the next authorized assistant can read.
• Instructions copied from outside material become project rules. Treat source content as material to evaluate, and keep approvals separate.

The guide to agent memory security covers access and untrusted content in more detail. For a Codex-specific setup, see Codex agent memory.

How to do this with fenbs

1. Create a free account and add the project whose work needs to continue.
2. Add or update its tasks with the current plan, completed work, outstanding checks, and references. Record choices people made in Decisions and rules.
3. Connect supported AI apps from Integrations using the browser sign-in. MCP is a standard way for an AI app to use an outside tool.
4. Before stopping, ask the outgoing assistant to save task progress and Where we left off, including the next step and anything waiting on a person.
5. Ask the incoming assistant to read the project context, starting with Rules and Where we left off. Check its proposed starting action, then have it save progress when its session ends.

fenbs does not import your chat history. It fits reviewed project progress; a shared document may be sufficient for a single transfer, and building memory inside your own application requires developer tools.