Zettelkasten Agent Memory: Linked Notes an AI Can Use

To use a Zettelkasten for agent memory, keep one understandable idea per note, attach its source, and explain its links to other ideas. Give your AI assistant a small, reviewed selection to consult, while keeping instructions in a separate record. The notes become useful reference material when another session can tell what they mean and how much to trust them.

Tools like fenbs keep the accepted lessons, project decisions, and next steps those notes produce outside any one AI conversation. Your linked notebook can remain a place to think without every tentative idea becoming a rule for the work.

How do you start with a few useful notes?

1. Choose one live question, such as whether a product instruction needs an example. Collect only the notes relevant to answering it.
2. Write a short note for each distinct claim. Use a title that states the idea, a source location, your interpretation, and its limits.
3. Link each note to a relevant existing idea and explain the relationship in a sentence: supports, challenges, narrows, or gives an example.
4. Ask the assistant to compare these notes against the question. Require note identifiers beside its claims and leave unsupported conclusions open.
5. Review the answer yourself. Save accepted conclusions with the project, and preserve unanswered questions for the next session.

Start with a handful you can read in one sitting. A large collection does not help if the assistant cannot identify the relevant claim or you cannot trace its summary. This workflow adds linked reference notes to agent memory, the information an assistant retains or retrieves between actions.

What should each note contain?

Write enough for someone who did not attend the original conversation. A note called “examples” says little. “A concrete example may resolve an ambiguous instruction” provides a claim you can inspect and connect to other observations. Avoid splitting a sentence from the qualification that makes it accurate.

• Identifier and title: a stable reference and a meaningful statement.
• Source: the document, author, link, and page or section you actually checked.
• Your explanation: what the source supports, in your own words.
• Limits: what remains uncertain and when the idea might not apply.
• Connections: related note identifiers and why the connection matters.
• Review state: unreviewed, checked against the source, or superseded, with a review date.

These are writing conventions you choose, rather than a required database format. If you use Obsidian, its properties documentation↗ (opens in a new tab) describes fields such as text, links, and dates. You can use those for review details while leaving the actual reasoning in the note body.

For a hypothetical support project, a note might report that readers misunderstood a particular installation step in a reviewed test. It should identify that test and its conditions. Do not turn that observation into “all instructions need screenshots” without evidence or a person choosing that policy.

Which connections help an assistant reason?

Prefer a few explained connections over many unexplained ones. “Related to onboarding” does not say whether the note supports an approach, contradicts it, or simply uses the same word. Write the reason beside the link so the assistant has an argument to inspect rather than a trail of similar titles.

A useful connection might say: “This observation challenges N-014 because the experienced participants also needed help.” Another might say: “N-022 describes an exception when the instruction is used offline.” Keep contradictory notes visible until you understand whether their contexts differ.

Obsidian supports links to notes and headings through its internal linking features↗ (opens in a new tab). The mechanics are useful, but the relationship still needs your explanation. For using memory during coding sessions in a particular client, see Codex agent memory.

Add one short entry note for your current question. List the few source notes that matter, the unresolved disagreement, and what evidence would change your view. This gives the next assistant a starting route through the collection without pretending that route is a complete account of the topic.

How do you keep reference material from becoming instructions?

Label reference notes as material to analyze. Keep the actual working instructions in a separate, clearly named place. A quotation that says “ignore earlier guidance” may be something you are studying; it should never acquire authority because it appears inside a retrieved note.

This boundary matters because external material can contain attempts to redirect an assistant, often called prompt injection. The NIST generative AI profile↗ (opens in a new tab) discusses that risk. Labels and careful prompts help communicate intent, but do not replace access controls or review before consequential actions.

Limit what the assistant can change during a research pass. Ask for suggested note edits first, then inspect the sources and approve the useful changes. Never give a notebook more sensitive information than the assistant and its service are allowed to handle. Our agent memory security guide covers that broader boundary.

When should a note become a lesson or decision?

A note explores an idea. A lesson describes what happened, what you learned, and what to do differently. A decision records what a person chose for the work. These records can point to the same evidence while doing different jobs.

Suppose the team reviews the installation test and chooses to add an example to one confusing step. Keep the observations in the notebook. Record the chosen change with its reason in the project, then add a task to revise and retest the instruction. A broader rule needs its own explicit human choice.

fenbs keeps those accepted outcomes in Decisions and rules, Lessons learned, and tasks; see how the work record fits together. That gives the next assistant a clear distinction between evidence worth reading and instructions people have agreed.

How to do this with fenbs

1. Create a free account and add the project your linked notes are helping you investigate.
2. Add useful source references to Links and listings and save a short reviewed context note explaining the relevant findings and uncertainties.
3. Record the human choice in Decisions and rules. Add a lesson when the work reveals a mistake to avoid, and create a task for the next research or implementation step.
4. Connect a supported assistant through Integrations. Ask it to read the project context before working and save progress, including unresolved questions, before it stops.

fenbs does not automatically import your linked notebook or provide a visual note graph. Keep using your note app when exploratory writing and following connections are the main job. Use the project record for the reviewed knowledge that changes what happens next.

Test the arrangement with a fresh session. Can the assistant identify the accepted decision, name the evidence behind it, and leave the unresolved interpretation open? If it cannot, simplify the project summary before adding more notes. The second-brain library offers other ways to make saved knowledge useful.