n8n-nodes-sql-chat-memory

Read-only AI Chat Memory for n8n. Loads conversation history from PostgreSQL — either by pointing at a table in the UI or with your own SQL — and hands it to the AI Agent as ordered chat messages.

What it does

Table/SQL → rows → LangChain chat messages → AI Agent memory

• A pure AI sub-node: connect its Memory output to an AI Agent's Memory input. It receives its data context (e.g. previous nodes' items for expressions like {{ $json.sessionId }}) through that connection — no Main input/output.
• Uses the standard n8n Postgres credential.
• Runs during AI Agent execution; converts rows in order:
  • user → HumanMessage
  • assistant → AIMessage
  • system → SystemMessage
• Strictly read-only: never saves, updates, inserts, caches, or auto-persists. No polling, session management, or schema creation.

Table Mapping mode (direct mapping, no SQL)

Six fields, top to bottom:

1. Schema / 2. Table — pick from the lists, or type a name.
2. Session Column + Session ID — the column that groups messages into one conversation, and the value for the current one (or an expression like {{ $json.sessionId }}). Leave the column empty if the table holds a single conversation.
3. Order Column — the date or number column that puts messages in order. Leave empty to detect it automatically. The newest messages are always loaded.
4. Role Column (required) + Role Mappings — the column that says who wrote each message, with every value matched to the person (user), the AI (assistant) or an instruction (system). Unmatched values stop the run and name the value — nothing is guessed.
5. Max Messages — how many past messages to load (1–1000, default 50).

The lists only load while you set up the node — never during a run.

Custom Query mode (advanced)

Only the SQL Query field. Write whatever you need — filtering, JOINs, message text built in SQL — as long as each result row looks like this:

[
  { "role": "user", "content": "hello" },
  { "role": "assistant", "content": "Hi, how can I help?" }
]

role is one of user, assistant or system. Load the newest X messages, but return them oldest first. You can use expressions. Read-only: SELECT queries only.

Missing role/content or an unsupported role produces a clear node error.

Custom query format

The Custom Query page ends with a guidance box showing this same JSON output format. Write the query however you want around that.

Intended workflow

Incoming message
    ↓
SQL Chat Memory ──────┐
                      ↓
                   AI Agent
                      ↓
                 AI response
                      ↓
             normal workflow nodes
                      ↓
             application-controlled
             database INSERT

Persist new turns with your own downstream INSERT. This node only reads.

Debugging

After an AI Agent run, open the SQL Chat Memory sub-node (or the agent's execution log): the exact messages handed to the AI are recorded there as { role, content } rows — the same shape as the Custom Query output. The node stays a pure memory sub-node with no Main input/output.

Versions

• v1 (npm 0.1.0): SQL-only sub-node, frozen unchanged.
• v1.1 (npm 0.2.x–0.4.5): Table Mapping UI plus Custom Query. Existing v1 workflows keep working untouched. Any table/column naming works — identifiers are quoted, never judged against examples.

Development

npm install
npm run build
npm test

Releasing (publish to npm via GitHub Actions)

1. Push to GitHub (main branch).
2. Repo Settings → Secrets and variables → Actions holds NPM_TOKEN (npm Automation token).
3. Bump version in package.json for every release (npm rejects re-publishing a version).
4. Create a GitHub Release (tag like v0.4.0). Publishing runs automatically (npm publish --access public).
5. Install in n8n via Settings → Community nodes → Install: n8n-nodes-sql-chat-memory.