Memory Sources

Every long-term memory the server extracts can say where it came from: the session and the events, or the document revision, that produced it. This answers "why do you believe that?" for an agent and "which conversation produced this wrong fact?" for an operator, and it makes "forget everything from this thread" a two-step workflow. Corrections, audit and memory poisoning are rising concerns in the n8n community and in other memory vendors' issue trackers; no other n8n memory node exposes provenance.

Include Sources

Search, Get and Get Many have an Include Sources switch (shown when Simplify is on). Each memory then carries sources, an array of:

Field	Meaning
sourceId	ID of the source link.
kind	SESSION or DOCUMENT.
sessionId	For a session source, the session the memory was extracted from.
documentId	For a document source, the ingested document.
eventIds	For a session source, the events that produced the memory.

It costs no extra request: the data is already in the response. This is the cheapest way to let a tool-using agent cite where a memory came from: "you told the support bot this on 12 May".

Get Sources

Long-Term Memory → Get Sources lists the sources of one Memory ID, one item per source, with more detail: revision for documents, eventIds across every evidence entry, evidenceCount (the number of evidence entries, not of events) and capturedAt (when the first evidence entry was recorded). With Simplify off, the raw link includes the evidence arrays and the producer that recorded them (iris-memory-extraction for the server's extractor).

A memory created directly with Create has no sources: nothing extracted it.

From a source to the evidence

Combine three operations to show the exact messages behind a memory:

flowchart LR
    S[Search<br/>Include Sources] --> G[Get Sources]
    G --> F{kind = SESSION}
    F --> R[Session → Get<br/>Include Summarized Events]
    R --> E[Filter events<br/>by eventIds]

Session → Get with Include Summarized Events on returns every event, including the older ones the summary already covers, so the evidence is found even after the session was summarized. The trace example wraps this in a webhook.

Correct a memory

A wrong or outdated fact ("lives in Berlin" after a move to Madrid) is fixed with Long-Term Memory → Update and a new Text. The memory keeps its ID and sources. Deleting it instead and letting the next conversation re-extract is also fine; the choice depends on whether the old sources still matter for audit.

Forget a thread

To remove a conversation and what was learned from it, for example one that came from an untrusted channel:

1. Session → Delete the session.
2. Long-Term Memory → Get Many with the Session ID and User ID filters, Return All.
3. Long-Term Memory → Delete each result.

Deleting the session alone does not delete the memories extracted from it: they are the user's long-term facts and stay until deleted. The same three steps, with the User ID filter alone and Session → Get Many by owner, implement "forget me".