AI Memory Graph: Connect Facts Without Losing Their Sources

An AI memory graph is a record that connects things, such as people, projects, decisions, and documents, through named relationships. An assistant can use those connections to find relevant information, but a useful graph must also show where each claim came from and whether it still applies. A line between two dots is not evidence on its own.

fenbs keeps approved work decisions and project context rather than providing a graph engine. That distinction matters when choosing between analyzing relationships and maintaining the facts that guide tomorrow’s work. Find the wider concepts in the AI memory library.

What do the dots and connections actually mean?

A dot, usually called a node, represents something identifiable: a project, a document, a person, or a decision. A connection, often called an edge, states a relationship between two things. Its label matters. “Authored,” “approved,” and “mentioned” describe different relationships, even if they connect the same person and document.

The W3C RDF primer↗ (opens in a new tab) describes one standard approach as a subject, a relationship, and an object. In plain words: “Decision A replaces Decision B.” Not every memory graph uses RDF, but the structure illustrates why a precise relationship is more useful than a vague association.

Also distinguish the stored information from its display. An AI memory map might be a drawing someone arranged manually. A graph-based record might support questions about connected facts without showing a visual map at all. Ask what the system stores and retrieves, rather than choosing based on how impressive the picture looks.

What would a useful project example look like?

Imagine a studio revising a restaurant website. The project has a booking form. A person approves a shorter form in a recorded decision, and a task implements that choice. A delivery test then checks whether the submitted message reaches the intended inbox. Those are related events, but they are not interchangeable proof of completion.

This structure helps answer “Which tasks depend on this approval?” and “What evidence supports calling this work complete?” It becomes misleading if the system changes “a design was discussed” into “a design was approved.” Record the actual person’s choice before giving the relationship that meaning.

For colleagues using different assistants, AI memory for teams addresses the shared record around these facts. A common vocabulary also helps: decide whether “form,” “booking page,” and “reservation request” refer to the same thing before joining their records.

Why should every important relationship keep a source?

An assistant may infer that two records are connected because they use similar language. That can help find candidates, but similarity does not establish identity, approval, or cause. Keep a proposed connection separate from a relationship directly supported by a source. Let someone review consequential claims before other work relies on them.

Provenance means the record of where information came from. The W3C PROV overview↗ (opens in a new tab) describes how sources, activities, and people can help assess data’s reliability. For project memory, useful source details include the original record, who made the claim, and when it was captured.

• Source: Which document, task, test result, or recorded conversation supports the connection?
• Scope: Which project, version, environment, or period does it concern?
• Status: Is the relationship confirmed, proposed, disputed, or replaced?
• Access: Can the person or assistant asking the question inspect the supporting evidence?

A graph answer should expose enough of this trail to check it. If the supporting source is missing or inaccessible, the answer should preserve that uncertainty. In fenbs, Decisions and rules record what a person chose; an assistant’s inferred connection does not become a new human decision.

How should the graph handle facts that change?

Record both when something was entered and when it was meant to apply. Those dates can differ. A person might record an approval on Thursday that takes effect the following Monday. A question about Friday’s work needs a different answer from a question about next Tuesday’s work.

A corrected error and a changed decision also need different treatment. If the graph assigned an approval to the wrong project, correct the error and preserve the reason for that correction. If a person deliberately chooses a new design, keep the earlier choice as history and connect the new decision to the one it replaces.

Define what “current” means before asking an assistant to choose between conflicting records. A newer capture date does not automatically outrank an older authoritative source. You may need to review the source, the decision maker, and the effective date together. Avoid silently deleting the older connection when its history explains existing work.

When is a graph worth the extra structure?

A graph is useful when relationships are the question: which projects share a dependency, which documents support a claim, or which tasks a replaced approval affects. It needs stable identifiers, clear relationship meanings, and maintenance. If the only recurring question is “Where did we stop?”, a short current progress record may answer it more directly.

Test with real questions before expanding. Ask for the source behind one answer and deliberately introduce an outdated sample decision. Can the system distinguish the old choice from the current one? Can you correct a mistaken connection without rebuilding the record by hand? Those checks matter more than the number of visible connections.

The separate AI agent knowledge-graph guide covers the agent-building intent. For carrying a deliberate work record between apps, see portable AI memory. Both benefit from clear sources, although they solve different practical problems.

How to do this with fenbs

1. Create a free account, then add one project and the tasks whose decisions you need to preserve. One person can start free with no credit card.
2. Record the person’s approved choices in Decisions and rules. Include the reason, scope, and source references; keep unresolved questions open.
3. Add context notes for supporting facts and Lessons learned for corrections that should guide future work. Keep original documents in their appropriate home.
4. Connect a supported AI app from Integrations. Ask it to read the project context, check relevant decisions, and save progress before it stops.

fenbs does not provide graph visualization, graph queries, or automatic extraction of relationships from your chats. Choose a graph system when that is the work you need to do. The AI brain introduction explains the broader purpose of maintaining useful work context outside any one assistant.