Create Memory

Estimated reading: 5 minutes

The Create Memory feature sets up a persistent, flow-level memory store that retains conversation data across sessions. It preserves past interactions and makes them searchable in future runs, so users do not have to repeat themselves.

Each memory store uses a configured embedding model and vector database to save conversations as vector embeddings. When connected to the Memory Base component in the same flow, workflows and AI agents can retrieve relevant past context, user preferences, and prior decisions through semantic search.

Note: A memory store belongs to the flow in which it is created. It can be accessed only through the Memory Base component within that flow and cannot be shared with other flows.

Prerequisites

Before creating a memory store, ensure your project has:

a. An Embedding Model configured and available.
b. A DB Provider configured and activated. (Only one DB Provider can be active per project at a time.)
c. An accessible Vector Database associated with your active DB Provider.

How Memory Storage Works

When a user interacts with your flow, conversational information is processed and stored in the following sequence:

1. Collection: User–AI message pairs are collected in a queue during the conversation.
2. Batching: Pairs are grouped according to your configured Batch Size. Ingestion begins only once the batch is full.
3. Preprocessing (Optional): If LLM Preprocessing is enabled, the selected language model summarizes or extracts key information from the batch using your instructions. Otherwise, raw messages are passed forward directly.
4. Embedding Generation: The configured embedding model converts the resulting content into vector embeddings.
5. Storage: The generated embeddings are saved to the selected vector database.

Memory Retrieval

When a workflow or AI agent needs context from past interactions, the Memory Base component performs a semantic search against the flow’s memory store. It evaluates the meaning and context of the incoming query to retrieve the most relevant historical records and inject them into the active run.

(For detailed configuration options, refer to the Memory Base Component documentation.)

Step 1: Create a Memory Store

1. Navigate to Login Robility Flow → Canvas → Right Panel → Memory icon (half-brain icon).
2. Select Create Memory.
3. Enter the required Basic Configuration details (see below).
4. (Optional) Enable and configure LLM Preprocessing.
5. Select Save Memory.

Basic Configuration

Field Description
Name* A unique name to identify the memory store.
Embedding Model* The model used to convert conversational content into vector embeddings for semantic retrieval.
Warning: This configuration cannot be changed after the memory store is created.
Vector Database* The database used to store and retrieve memory embeddings. Only databases associated with the project's active DB Provider appear in the dropdown.
(See DB Providers documentation [AB2.1] for setup.)
Batch Size* The number of user–AI message pairs (1 User Message + 1 AI Response = 1 Pair) collected before memory ingestion begins. Refer Batch size details.

Choosing a Batch Size

Processing starts only when the configured batch is full. Select a batch size based on your expected conversation volume and how quickly new information needs to become searchable:

Batch Size Recommended Usage Processing Behavior & Trade-offs
1 Low-volume conversations requiring immediate recall Processes each user–AI message pair individually as it occurs; highest processing frequency.
5 Moderate conversation volumes Processes five user–AI message pairs together; balances latency and resource usage.
10 or higher High-volume conversations where immediate availability isn't critical Collects more message pairs before processing, reducing processing frequency but increasing the delay before information becomes retrievable.

Step 2: Configure LLM Preprocessing (Optional)

By default, original conversational messages are embedded directly. Enabling LLM Preprocessing uses a language model to summarize conversations or extract specific facts before generating embeddings.

Note: Enabling LLM preprocessing introduces additional model usage costs and processing latency for every batch. Enable it only when summarization or noise filtering is required.

Field Description
LLM Preprocessing Enable this option to process conversational content with a language model before embedding generation.
Preprocessing Model The language model used to condense or extract information. (See Model Providers documentation for setup.)
Preprocessing Instructions Prompt instructions guiding the model on which details to retain or exclude.

Example Preprocessing Instructions

a. Summarize Conversation: “Summarize the conversation, retaining important decisions, action items, and user preferences.” 
b. Extract Key Information: “Extract important facts, preferences, and action items while excluding greetings and unnecessary conversational content.”
c. Capture Business Context: “Retain project requirements, decisions, blockers, and assigned tasks.”
d. Customer Support Context: “Retain customer issues, resolutions, product references, and follow-up actions.”

Step 3: Attach a Memory Base Component to a Flow

Creating a memory store does not automatically connect it to your workflow. You must add and configure the Memory Base component to access the stored memories during flow execution:

1. Open the flow in Robility Flow where the memory store was created.
2. Add the Memory Base component to the canvas.
3. Open the component configuration panel and select your memory store from the Memory Base dropdown.
4. Configure any remaining component parameters and connect it to the appropriate workflow or AI agent.
5. Save the flow.

Limitations

a. Immutable Embedding Model: The embedding model configuration cannot be modified after a memory store is created. To use a different model, you must create a new memory store.
b. Preprocessing Requirements: When LLM preprocessing is enabled, both a preprocessing model and preprocessing instructions are required.
c. Ingestion Delay: Memory ingestion begins only when the configured Batch Size threshold is reached.
d. Performance Dependencies: Retrieval accuracy depends on the chosen embedding model and the quality of stored content, while storage capacity and query performance depend on the selected vector database.

Share this Doc

Create Memory

Or copy link

CONTENTS
Robility Chatbot
Robility Assistant
Online