Reverts the removal in 2a61fe1a. That commit dropped app_id, includes and
excludes from Add because docs/openapi.json does not declare them on
POST /v3/memories/add/. The schema is incomplete: probing the live endpoint
shows all three are accepted and honoured.
- app_id is persisted and is queryable via /v3/memories/search/ and
/v3/memories/, so the entity-id guard accepts it again as a valid scope
- includes/excludes demonstrably steer extraction. The same message stored
three memories unfiltered and one under
includes: "only record food and diet preferences"
Search and Get Many gain an App ID filter alongside User/Agent/Run. Writing
an app-scoped memory was otherwise unreadable from this node, and app_id is
how the OpenCode and Pi Agent plugins express project scope.
buildEntityFilters now takes the ids as an object rather than four adjacent
positional strings, so a fifth entity is a one-key change.
Three review findings on the Mem0 n8n node:
- Memory IDs and event IDs were interpolated raw into request paths, so a
value containing / or .. could reach a different endpoint than intended.
Wrap all four sites in encodeURIComponent.
- Search and Get Many could only filter by user_id, leaving agent- and
run-scoped memories unreachable from the node. Both now accept User ID,
Agent ID, and Run ID. Multiple ids combine with OR because Mem0 indexes
entities separately and an AND across them matches nothing. User ID is
no longer marked required; the node still fails with a clear message
when all three are empty, rather than letting the API return a raw 400.
- n8n-nodes-mem0-checks.yml ran lint and build but never the jest suite,
so the tests could rot unnoticed. Add a test job matching the layout of
pi-agent-plugin-checks.yml.
Tests cover flat single-entity filters, the OR combination, agent-only
listing, the empty-entity guard on both operations, and id escaping in
the get/delete/poll URLs. Docs and README document the new fields and the
OR semantics.
The Add operation exposed app_id, includes, and excludes in Additional
Fields and forwarded them to POST /v3/memories/add/. None of the three
are declared on that endpoint's schema (they belong to MemoryInput, used
by POST /v1/memories/), so they were silently dropped by the server.
app_id was worse than a no-op: the entity-id guard accepted it as
satisfying "at least one entity id", so filling only App ID passed local
validation and then sent a v3 request carrying no scope v3 recognises.
The guard now requires User ID, Agent ID, or Run ID.
Docs corrections alongside it:
- Add: document Custom Instructions and Custom Categories, which the
node has always supported but the docs never mentioned
- Get Many: document Return All, which the docs omitted entirely
- README: Infer and Wait for Completion are independent controls; the
README claimed infer=false returns synchronously, contradicting the
node's own field description and its actual behaviour
Every other publishable integration (@mem0/pi-agent-plugin,
@mem0/openclaw-mem0, @mem0/vercel-ai-provider) and the repo root ship
Apache-2.0. The n8n node carried MIT from the n8n starter template.
LICENSE is now byte-identical to integrations/pi-agent-plugin/LICENSE.
eslint-plugin-n8n-nodes-base hard-codes MIT in
community-package-json-license-not-default, so that rule is disabled.
It is a starter-template convention, not a functional requirement: n8n
resolves and loads community packages without reading the license field,
and n8n's own node packages (n8n-nodes-base, @n8n/n8n-nodes-langchain)
publish as "SEE LICENSE IN LICENSE.md" rather than MIT.
The n8n node was the only integration without a `pnpm.overrides` block,
so its tree carried 9 advisories (7 high, 2 moderate) that
pi-agent-plugin, openclaw, and vercel-ai-sdk already pin out. Adds the
same block, scoped to the packages actually present in this tree
(`esbuild` and `undici` are not, so their sibling entries are omitted
rather than carried over as dead config).
Down to 2 from 9. Both remaining are GHSA-mh99-v99m-4gvg on the 1.x and
2.x brace-expansion lines, whose fix shipped only in 5.0.8 with no
backport. Forcing those lines to 5.x does clear the audit but breaks
`gulp build`: rimraf -> glob -> minimatch@9 imports brace-expansion as a
default export, which the 5.x ESM build does not provide.
Replacement ranges are capped per major rather than left open like the
siblings'. An unbounded `uuid >=11.1.1` resolves to 14.x, which is
pure ESM and cannot be required by n8n-workflow under Jest's CJS
runtime.
All dev-only: the published package still declares zero runtime
dependencies, so none of this reaches an installed node. lint, build,
and the 5 tests pass; `pnpm install --frozen-lockfile` is clean.
The unscoped `n8n-nodes-mem0` name is already taken on npm by an
unrelated publisher, so the CD workflow could never publish it. Move to
the `@mem0` scope, matching `@mem0/pi-agent-plugin` and
`@mem0/openclaw-mem0`, and add `publishConfig.access: public` since
scoped packages default to restricted.
n8n accepts scoped community packages: it strips the scope before
enforcing the `n8n-nodes-` prefix, the same shape as n8n's own
`@n8n/n8n-nodes-langchain`. Verified end to end on n8n 2.22.6 by
installing the packed tarball into the community-packages directory and
executing a workflow against node type `@mem0/n8n-nodes-mem0.mem0`.
The directory, workflow filenames, and the `n8n-nodes-mem0-v*` tag
prefix are unchanged. pnpm-lock.yaml needs no update: the pnpm v9 root
importer is anonymous and never recorded the package name.
kartik-mem0