AI Agent Tools for n8n

This package includes an AI Agent Tool node that allows n8n’s AI Agents to interact with your vector store using intuitive operations.

---

What is the PGVector Store Tool?

The PGVector Store Tool provides AI Agents with a natural way to manage knowledge:

Operation	What it does
Remember	Store new information, or update by ID or by finding similar content
Recall	Search for similar information with configurable threshold
Forget	Delete a specific entry by its exact ID
Forget Similar	Delete entries similar to a concept (with safety controls)
Lookup	Get a specific entry by ID

---

Quick Start

1. Add the AI Agent

1. Create a new workflow in n8n
2. Add an AI Agent node
3. Connect a Chat Model (e.g., OpenAI, Anthropic)

2. Add the PGVector Store Tool

1. Add a PGVector Store Tool node
2. Connect it to the AI Agent’s tool input
3. Connect an Embeddings node (e.g., OpenAI Embeddings)

3. Configure the Tool

1. Set up your Postgres credentials
2. Choose the Operation
3. Configure operation-specific settings
4. Specify your Collection name

---

Operations

Remember (Store/Update)

Store new information or update existing entries.

n8n Configuration: | Setting | Description | |———|————-| | Collection | Knowledge base name | | ID Format Hint | Example ID format to guide the AI (e.g., “meeting-2024-01-15”) | | Auto-Generate ID | Automatically create IDs when AI doesn’t provide one | | Update Similarity Threshold | Min similarity to allow concept-based update (default: 0.7) | | Distance Metric | For concept-based updates (cosine, l2, inner_product) |

AI provides: | Parameter | Description | |———–|————-| | content | The information to store (required) | | id | Update entry with this ID | | updateSimilar | Find and update entry similar to this concept | | metadata | Tags like {category: "meeting"} |

Example AI interactions:

Create new:

“Remember that the API key expires on January 15th”

Create with specific ID:

“Remember with ID ‘api-expiry’: The API key expires January 15th”

Update by ID:

“Update entry ‘api-info’ with the new expiration date”

Update by concept:

“Update the information about API keys with this new content: …”

How it works in n8n:

1. Configure ID Format Hint to “api-v1-info” so the AI uses consistent IDs
2. Enable Auto-Generate ID if you don’t need specific IDs
3. Set Update Similarity Threshold to 0.8 for strict concept matching

Recall (Search)

Search for similar information.

n8n Configuration: | Setting | Description | |———|————-| | Collection | Knowledge base name | | Top K Results | Maximum results to return (default: 5) | | Minimum Similarity | Only return results above this threshold (0-1, default: 0) | | Distance Metric | cosine, l2, or inner_product |

AI provides: | Parameter | Description | |———–|————-| | query | What to search for (required) | | filter | Metadata filter like {category: "meeting"} |

Example AI interactions:

“What do we know about the API?”

“Find all meeting notes from Q1”

How it works in n8n:

1. Set Minimum Similarity to 0.6 to filter out weak matches
2. Use Top K Results = 3 for focused responses, 10 for comprehensive

Forget (Delete by ID)

Delete a specific entry by its exact ID.

n8n Configuration: | Setting | Description | |———|————-| | Collection | Knowledge base name | | ID Format Hint | Example ID format so AI knows what IDs look like | | Return Deleted Content | Show what was deleted (useful for confirmation) |

AI provides: | Parameter | Description | |———–|————-| | id | The exact ID to delete (required) |

Example AI interactions:

“Delete the entry with ID ‘old-notes-123’”

How it works in n8n:

1. Set ID Format Hint to match your ID scheme (e.g., “doc-123”)
2. Enable Return Deleted Content for audit trails or confirmations

Forget Similar (Delete by Concept)

Delete entries similar to a concept. Has safety controls.

n8n Configuration: | Setting | Description | |———|————-| | Collection | Knowledge base name | | Distance Metric | For similarity matching | | Similarity Threshold | Only delete if similarity above this (default: 0.8) | | Dry Run | Show what would be deleted without deleting (default: ON) |

AI provides: | Parameter | Description | |———–|————-| | concept | Delete entries similar to this (required) |

Example AI interactions:

“Delete all information about the deprecated API”

How it works in n8n:

1. Keep Dry Run ON for safety - shows what would be deleted
2. Set Similarity Threshold to 0.9 for very precise matching
3. Create a separate workflow with Dry Run OFF for actual deletion

Lookup (Get by ID)

Retrieve a specific entry by its exact ID.

n8n Configuration: | Setting | Description | |———|————-| | Collection | Knowledge base name | | ID Format Hint | Example ID format so AI knows what IDs look like | | Include Metadata | Show metadata tags in response (default: ON) | | Include Timestamps | Show created/updated times (default: ON) |

AI provides: | Parameter | Description | |———–|————-| | id | The ID to retrieve (required) |

Example AI interactions:

“Show me the entry with ID ‘meeting-notes-jan’”

How it works in n8n:

1. Set ID Format Hint to help AI recognize valid IDs
2. Disable Include Timestamps if you don’t need date info

---

Tool Naming

Tools are automatically named based on collection:

• remember_knowledge
• recall_knowledge
• forget_knowledge
• forget_similar_knowledge
• lookup_knowledge

---

Configuration vs AI Parameters

The design philosophy separates concerns:

Controlled by	Purpose	Examples
n8n Configuration	Safety, limits, behavior	Thresholds, dry run, ID hints
AI Runtime	Data and targets	Content, IDs, search queries

This means:

• Users control safety (thresholds, dry run mode) via n8n UI
• AI handles the data (what to store, what to search for)
• ID Format Hints guide the AI without enforcing - they appear in tool descriptions

---

Example Workflows

Knowledge Base with Full CRUD

Chat Trigger → AI Agent → OpenAI Chat Model
                    ↓
              PGVector Store Tool (Remember) → OpenAI Embeddings
              PGVector Store Tool (Recall) → OpenAI Embeddings
              PGVector Store Tool (Forget) → OpenAI Embeddings
              PGVector Store Tool (Lookup) → OpenAI Embeddings

Configuration:

• Remember: ID Format Hint = “note-YYYY-MM-DD”, Auto-Generate ID = ON
• Recall: Top K = 5, Minimum Similarity = 0.5
• Forget: ID Format Hint = “note-YYYY-MM-DD”, Return Deleted Content = ON
• Lookup: Include Metadata = ON, Include Timestamps = ON

Safe Cleanup Workflow

Chat Trigger → AI Agent → OpenAI Chat Model
                    ↓
              PGVector Store Tool (Forget Similar, Dry Run ON)

First run with Dry Run to see what would be deleted, then create another workflow with Dry Run OFF for actual deletion.

Meeting Notes Assistant

Chat Trigger → AI Agent → OpenAI Chat Model
                    ↓
              PGVector Store Tool (Remember) → OpenAI Embeddings
                    ↓
              PGVector Store Tool (Recall) → OpenAI Embeddings

Configuration:

• Remember:
  • ID Format Hint = “meeting-2024-01-15-standup”
  • Auto-Generate ID = OFF (AI provides structured IDs)
  • Update Similarity Threshold = 0.85
• Recall:
  • Top K = 10
  • Minimum Similarity = 0.6

User interaction:

User: “Remember today’s standup: John discussed the API migration” AI: Uses remember_meetings with ID “meeting-2024-01-15-standup”

User: “What did we discuss about APIs last week?” AI: Uses recall_meetings to search

---

Best Practices

1. Use ID Format Hints

Set the ID Format Hint in n8n to guide the AI:

meeting-2024-01-15-standup
doc-api-v2-guide
note-project-alpha-123

The hint appears in the tool description, helping the AI use consistent IDs.

2. Enable Auto-Generate ID for Simple Cases

When you don’t need specific IDs, enable Auto-Generate ID:

• Creates IDs like knowledge-1705315200000-x7k2m9
• Prevents missing IDs when AI forgets to provide one

3. Use Concept Updates Carefully

The updateSimilar feature is powerful but should have a reasonable threshold:

• 0.85+ for strict matching (production)
• 0.7 for moderate matching (development)
• Below 0.6 may match unintended entries

4. Always Dry Run First

When using Forget Similar, always test with Dry Run ON to see what would be deleted.

5. Return Deleted Content for Auditing

Enable “Return Deleted Content” on Forget to:

• Confirm the right entry was deleted
• Provide audit trail for users

6. Organize with Metadata

{
  "metadata": {
    "category": "meeting",
    "quarter": "Q1",
    "project": "alpha"
  }
}

Use with Recall’s filter parameter to narrow searches.

---

Troubleshooting

“No existing entry found similar to…”

When using updateSimilar, no entry was found that matches. Store as new instead.

“Similarity below threshold”

The found entry isn’t similar enough to the search concept. Either:

• Use the provided ID to force update
• Lower the Update Similarity Threshold in n8n config

“Dry run - would delete X entries”

This is expected! Disable Dry Run in n8n config to actually delete.

AI doesn’t use the ID format I specified

ID Format Hint is guidance, not enforcement. To ensure consistent IDs:

1. Make the hint clearer (e.g., “project-YYYYMMDD-topic”)
2. Include format in your prompts to the AI
3. Or enable Auto-Generate ID for automatic IDs

---

See Also

• Installation Guide
• Operations Reference
• MCP Server - Use with external AI agents