Kiro Agent Memory: Keep Useful Context Across Sessions

Kiro agent memory combines learned context with explicit steering, the project guidance you maintain in files. To make important context survive a new session, put mandatory guidance in steering, inspect any learned memory you rely on, and test both in a fresh conversation.

A separate record such as fenbs can hold shared project decisions and handoffs when your work also involves other assistants. fenbs does not have a listed Kiro setup guide. Use a supported assistant or manual review to bring the relevant shared record into your Kiro work; do not assume an automatic connection.

How should you start without saving everything?

1. Choose one repository and one recurring problem. For example, Kiro keeps proposing a new test library when the project already has a preferred approach.
2. Write the approved rule in workspace steering. Include the desired behavior and a useful example from the repository, then review the wording yourself.
3. Check the Memory controls for the Kiro app you use. Decide whether learned memories should be read, changed, or left off.
4. Inspect the saved guidance and start a fresh session. Ask Kiro what it would do for a small related task before permitting edits.
5. Compare its answer with your rule. Correct the source or its loading conditions before adding a larger collection of context.

A narrow test is easier to judge than “remember our whole project.” It also makes wrong assumptions visible early. If the assistant chooses the right library for the wrong reason, ask it to identify the guidance it used and inspect that evidence before calling the setup reliable.

Which information belongs in steering or learned memory?

Kiro’s memory documentation↗ (opens in a new tab) distinguishes learned memories from steering and session history. Learned memory can carry useful context into later work; steering is where you explicitly state guidance that must apply consistently. Neither should be treated as a complete copy of every past conversation.

Say a developer discovers that one test needs a prepared sample file. A reusable note should explain that prerequisite and point to the maintained instructions. A temporary error dump belongs with the investigation. Copying the whole dump into every future session makes the useful sentence harder to find.

The distinction is the same one behind agent memory versus context: some information guides this conversation, while a smaller reviewed set deserves a continuing home. Keep that set understandable to a teammate who was not present.

How do you make project guidance available deliberately?

Kiro steering↗ (opens in a new tab) uses Markdown files, ordinary text documents with simple formatting. Workspace steering lives in the project’s .kiro/steering/ directory. Scope repository-specific guidance there rather than turning one project’s exception into a preference for every project.

Give each instruction a concrete trigger and outcome. “Use the existing contact-form validation helper for new form fields” is more useful than “follow best practices.” Explain the reason when it changes judgment, such as preserving the same error messages across the website. Link to a representative file rather than reproducing its entire implementation.

Custom agents need an extra check: Kiro’s steering guide says their steering files must be included explicitly through the agent’s resources. If guidance works with one agent and disappears with another, inspect that configuration before rewriting the rule. Also check whether the instruction’s inclusion conditions match the files being changed.

Keep project decisions distinct from local coding guidance. A team may decide that a feature needs client approval while the repository explains how to test it. In fenbs, the shared project context can preserve the approval requirement and current work, while steering points Kiro toward the relevant implementation conventions.

How can you prove what survives a new session?

Test a rule and a fact separately. For the rule, ask Kiro to describe its approach to adding one form field. For the fact, save a harmless, distinctive detail about a sample project and ask for it after restarting. A correct answer should identify the relevant saved source and explain the consequence for the work.

Check learning and recall separately too. Read access can allow an existing memory to be used without allowing new ones to be saved. Kiro’s Memory guide describes access controls and ways to inspect stored memories; follow the instructions for your app because its controls are not interchangeable with another Kiro surface.

For CLI context, Kiro documents `/context show` and temporary session additions↗ (opens in a new tab). Use the displayed context to distinguish persistent agent resources from files added only for the current session. If a file was temporary, its absence from a new chat is expected rather than proof of a broken save.

Record the test conditions: app, repository, selected agent, saved item, and exact question. Repeat it after a meaningful configuration change. Do not infer that every repository, account, or computer shares context just because one local restart succeeded.

What should a useful handoff contain?

A useful handoff tells the next session what changed, what was checked, what is uncertain, and where to begin. Include the decision that limits the next action. “Validation helper updated; error-message check still pending; client approval required before adding another field” gives much more direction than “continue the form work.”

Do not turn untested output into a durable success claim. If Kiro proposed a fix but nobody ran the relevant check, preserve that distinction. Another assistant should be able to continue the investigation without first discovering that “finished” meant “suggested.” The broader agent memory guide explains why useful memory needs selection and review.

How to do this with fenbs

1. Create a free account and add the project you are working on in Kiro. Write the current goal and the next task.
2. Record choices a person made in Decisions and rules. Add project facts as AI context notes and incomplete work as tasks with their remaining checks.
3. Choose a compatible assistant from Integrations, or manually review the relevant project context before bringing it into Kiro. fenbs does not provide a listed Kiro setup guide.
4. After the session, update the task and Where we left off with what changed, what was checked, and the next step. Compare that record with the handoff you give the next assistant.

fenbs is not the right fit if all you need is one repository’s static coding instructions. Steering files may cover that need on their own. It becomes useful when project work must remain understandable beyond one tool’s local context.

How do you keep this setup manageable?

Review a saved item when its underlying fact changes. Replace temporary exceptions, keep a clear current decision, and remove repeated explanations that no longer help the next action. Your memory system should reduce the amount people need to reconstruct, not create another document collection nobody trusts.

For a broader implementation approach, see how to build agent memory and the AI memory library; Codex agent memory covers that assistant’s separate setup. Begin here with one rule, one saved fact, and one restart test you can inspect.