Skip to main content
n8n-nodes-haki gives any n8n AI Agent workflow durable, cross-session memory on your own Postgres — no vendor lock-in, no per-call memory tax. Two nodes, placed around your existing AI Agent node: one reads memory in, one writes the turn back out.
Runs against any Haki instance you already host (docker compose up). No Haki account required — self-hosted end to end.

​
Install

Or from the CLI, inside your n8n data directory:

​
Credential — Haki API

One credential, shared by both nodes:
​
string
required
The Haki API URL, no trailing slash. From an n8n Docker container to a Haki API on the host machine, use http://host.docker.internal:8100 instead of localhost.
​
string
Haki bearer key. Optional in local development (the API runs in open mode), required as soon as HAKI_ADMIN_KEY or a project key is configured server-side.

​
Haki Context — place before your AI Agent

Fetches the subject’s memory as a ready-to-inject block, before the LLM call.
​
string
required
Haki project (memory scope). Never chosen by the model.
​
string
required
Stable identifier for the person (sessionId, email, webhook ID…). Empty or "default" is rejected.
​
string
required
The current message — used to rerank relevant facts.
​
number
default:"2000"
Token budget for the returned context packet.
​
string
Optional task type (support, onboarding…), recorded in the trace.
Outputs context_text (drop straight into your AI Agent’s system message), the raw packet, warnings, token_count and trace_id.

​
Haki Capture — place after your AI Agent

Records the user/assistant turn once the reply exists.
​
string
required
Must match the Haki Context node.
​
string
required
Reuse the one from the Haki Context node.
​
string
required
​
string
required
​
string
Optional — either anchors an idempotency key derived from the run/ thread plus a content hash, so re-running the same n8n execution never duplicates the captured turn.
​
boolean
default:"false"
When on, also calls /v1/consolidate so the memory is recallable immediately — useful in dev/demo, not needed in steady-state production (consolidation runs on its own cadence).
​
options
default:"trusted"
trusted (direct message from the subject), semi_trusted (agent/tool output), third_party (someone else in the conversation) or untrusted (ingested content). In a group where the sender is not the tracked subject, choose third_party — the fact is attributed to that third party, never silently merged into the subject’s own memory.

​
A minimal wiring

Haki Context runs first and hands context_text to the AI Agent’s system message; Haki Capture runs after the agent replies, so the next turn — even in a brand new execution — starts with everything the agent already learned.

Ready-made workflows

A persistent support-agent template and an end-to-end test workflow ship in the repository’s integrations/n8n/ directory — import either one directly into n8n to see the wiring above running for real.