n8n-nodes-itsuki
v0.2.0
Published
Itsuki memory for n8n workflows and AI Agents: save durable facts, recall them in any later run, and hand memory to an AI Agent as a tool.
Maintainers
Readme
n8n-nodes-itsuki
Long-term memory for n8n workflows and AI Agents, backed by Itsuki. Save durable facts in one run, recall them in any later run, and hand memory to an AI Agent as a tool it calls itself.
This is a community node for self-hosted n8n. n8n Cloud carries only nodes n8n has verified; until that happens, Cloud users can connect Itsuki with the built-in HTTP Request or MCP Client Tool nodes instead — see the Itsuki n8n guide.
Install
In n8n: Settings → Community Nodes → Install, enter n8n-nodes-itsuki, acknowledge the risk notice, install. Only instance owners can install community nodes.
Credential
Create an Itsuki API credential with a project key from the Itsuki dashboard (API Keys). The key travels only as an Authorization header — never in URLs, exports, or logs. The built-in credential test calls a content-free status endpoint. Base URL stays https://itsuki.app unless you run a development server (HTTPS required; plain HTTP only for localhost).
Operations
| Operation | What it does |
|---|---|
| Save Memory | Store one durable fact. Returns the real receipt. |
| Save Conversation | Store a batch of chat messages (≤30, order preserved) for extraction, with optional narrowing (last-N / topic / summary). |
| Recall Memory | Semantic recall: a prompt-ready context string plus structured items. Zero results is a result, not an error. |
| List Memories | One page, or Return All with bounded automatic pagination (hard cap, loop detection). |
| Get Memory | Fetch one memory by prefixed ID. |
| Delete Memory | Delete one memory by ID. Destructive; a missing ID is an error, never a silent success. |
| Delete All Memories | Previews by default. Destroys only with the explicit Confirm Destruction toggle. |
| Who Am I | Connection health, identity, and plan usage. Never memory content, never the key. |
| Wait for Packet | Poll a save receipt until extraction settles. |
Wait for Completion is on by default for both save operations: extraction is asynchronous, and waiting means the next node sees the settled state. On timeout the output says, honestly, that the save was accepted and still processing — a timeout is not a failure.
Use as an AI Agent tool
The node is usableAsTool. Attach one node set to Recall Memory and one set to Save Memory (or Save Conversation) to a Tools Agent; keep the User ID stable across both. On self-hosted n8n, community nodes additionally require the environment variable N8N_COMMUNITY_PACKAGES_ALLOW_TOOL_USAGE=true to appear as tools.
Destructive operations are deliberately unattractive as tools: the node defaults to Recall, Delete All previews unless a human enables Confirm Destruction, and nothing auto-escalates.
Scoping
Tenancy comes from the credential (one key = one project) and can never be widened by a request. Optional fields narrow attribution: User ID isolates one end user's memory space under your key (use the same stable value on save and recall), and Conversation/Thread/Source IDs record where a memory came from.
Deliberately absent, because the backend governs them at the project level rather than per call: per-request extraction instructions, custom categories, verbatim (no-inference) storage, and arbitrary metadata. What extraction keeps is project policy in Itsuki, not a per-call knob. There is also no Update operation yet — the backend has no safe versioned-correction contract, and this node does not fake one.
Reliability
Bounded timeouts; cancellation propagates to in-flight requests and backoff sleeps; Retry-After is honoured exactly; reads retry with jittered backoff; writes retry only when an Idempotency Key protects them; deletes never retry. Rate-limit (429), quota, and backlog responses map to named, friendly errors with no secrets in them.
License
Apache-2.0
