AI Memory Files: What to Save Between Sessions

AI memory files are documents an assistant reads to recover useful information from earlier work. Save durable instructions, verified project facts, and a short handoff that says what is finished, what remains uncertain, and where the next session should begin.

The files help only when the next assistant can find and read them. A local document can work well for one person in one workspace. When the work needs to reach several people or AI apps, fenbs offers a shared project record with context notes, Decisions and rules, and Where we left off.

How do you make file-based memory work?

1. Pick one project and list what its next session must know before acting.
2. Separate lasting instructions from facts and from the current handoff. Give each file a clear purpose.
3. Add sources, the person responsible for the record, and the date of the last meaningful review.
4. Configure your assistant’s supported instruction mechanism or explicitly ask it to read the files.
5. Start a fresh session and ask it to identify the current rule, the remaining check, and the next action.

Do this before copying a long chat transcript into a folder. A transcript contains proposals, abandoned ideas, corrections, and successful results mixed together. A memory file should explain what a future session can rely on, while pointing to the original evidence when it matters.

The AI memory topic hub covers other approaches. For files, the practical question is simple: can someone open the project tomorrow and tell what to do next without reconstructing yesterday’s conversation?

What belongs in each file?

Use a small structure that matches the work. The filenames below are suggestions for your own documents, not a convention every AI app automatically recognizes. If your assistant has a required instruction filename, keep that file as the entry point and direct it to supporting material.

You can begin with sections in one document instead of four files. Split a section when it has a different owner, changes at a different pace, or distracts from the next task. A folder full of nearly identical summaries is harder to trust than one short current record.

Give each important fact one current home. Other files can point to it. If the delivery test result changes, update the work-state record rather than copying the new result into every instruction file.

How do you write a handoff the next session can use?

Write the handoff around the next action, not the order of yesterday’s conversation. Say what changed and what was checked. Separate a completed edit from a verified outcome. “Form built; delivery untested” tells the next assistant something that “contact form done” hides.

Project: Harbor Workshop website
Current goal: Finish the approved contact form
Approved source: Design C, linked in PROJECT-FACTS.md
Completed: Three required fields implemented
Checked: Required-field validation works locally
Not checked: Message delivery
Next step: Submit a test message and inspect delivery
Waiting on: Project owner to approve publication
Last reviewed: October 20, 2026
Owner: Project lead

Include enough detail to resume, then link to larger artifacts. A report, design, or test log can stay in its normal location. The handoff needs to say which artifact is current and why it matters. Do not bury the next step beneath a full history of commands or drafts.

When work moves between assistants, the portable AI memory guide explains the broader handoff. In fenbs, Where we left off keeps that starting point with the project, so connected assistants can pick it up.

How do you make sure the assistant reads the right files?

Check the documentation for the app you actually use. Claude Code’s memory documentation↗ (opens in a new tab) explains persistent instruction files and its separate automatic memory notes. OpenAI’s AGENTS.md guide↗ (opens in a new tab) describes how Codex discovers project instructions and applies guidance from different directory levels.

Those mechanisms do not mean any document named “memory” will load everywhere. Place the instruction entry point where the app expects it, and keep a short pointer to your project facts and handoff. If the app cannot access the folder, provide the relevant text explicitly through a method it supports.

Test discovery separately from recall. Ask a fresh session which project instruction file it found and what handoff it read. Compare that answer with the files. Then ask it to perform a small task that depends on them. An assistant can repeat a filename without using its contents correctly.

Our comparison of AI context files covers the file conventions in more depth. Keep this first setup focused on the few documents your next session actually needs.

What should stay out of memory files?

Leave out passwords, tokens, customer records, and unrelated personal details. A convenient project file may travel through source control, backups, copied folders, or another person’s workspace. Save a safe reference to an approved source when necessary, rather than putting sensitive content into general instructions.

Do not turn guesses into rules. Mark an unverified finding as uncertain and say what would confirm it. If a person approved a requirement, identify that approval; if the assistant merely suggested a change, keep it a suggestion. File proximity does not make the two equivalent.

For shared work, decide who maintains the current answer and who can read it. AI memory for teams covers that responsibility. A file on your laptop cannot help a colleague’s assistant unless its workflow gives it access.

How to do this with fenbs

1. Create a free account and add the project represented by your files.
2. Review the files and copy only current, verified project facts into context notes. fenbs does not import your local memory files automatically.
3. Record people’s approved decisions and standing instructions in Decisions and rules, then add tasks for unfinished work.
4. Connect a supported assistant using Integrations. MCP is a standard way for an AI app to use an outside tool, with browser sign-in for fenbs.
5. Ask the assistant to read project context before working and save progress before stopping. Check a new session against the original handoff.

fenbs is free for one person, no credit card. It is not the right fit if you need an offline file that travels with a repository and no outside service. A small file-based setup can be enough while all the work stays in one accessible workspace.

How do you keep the files trustworthy after each session?

Before stopping, replace the old stopping point with the current one, record the checks actually run, and remove instructions that a person has superseded. Preserve useful historical evidence separately from the current answer. Ask the next session to confirm the remaining task before making changes.

For the general concept, read the AI memory explanation. A well-maintained memory file has a narrower job: it carries a small amount of correct, reachable information into the next session, with enough evidence for somebody to challenge it.