Memories Overview

Memories are documents stored in folders that your workflows can create, update, search, and retrieve. Unlike files attached to prompts, memories are dynamic - workflows modify them over time.

Memories are flexible. They work as:

• Persistent context - a customer profile or running research document that a workflow updates with each new interaction
• Workflow state - tracking what a scheduled workflow has already processed so it doesn't duplicate work on the next run
• Knowledge base - a collection of documents indexed for semantic search, used as RAG context in prompts
• Cross-workflow data - one workflow writes findings, another reads them; no external database required

Each memory has a name, markdown content, and optional metadata fields (key/value pairs) you define. Memories live in folders you create and can be shared across your organization.

Folders

Navigate to Tools > Memories to access your folders.

Create folders around specific purposes: "Customer Profiles", "Research Findings", "Meeting Notes", "API Responses". Once created, a folder is available to any workflow you give access to.

Each folder has a Settings page with three connectivity toggles and a vector description field:

Enable Knowledge Base Search - turns on vector embedding. When enabled, memories are indexed after creation or update and become searchable by semantic meaning. Required for vector search to work.

Enable Chat and Prompts - makes this folder attachable to prompts and usable as a knowledge base in Projects.

Enable External MCP Servers - exposes this folder via the Memories MCP server so external tools can connect to it with a bearer token. See Memories MCP Server.

Vector description - an optional description prepended to embeddings for the whole folder, helping the model understand what the folder contains. Can also be set per-document to override the folder-level description.

Creating Memories

Manually - click New Memory in any folder, give it a name, write the markdown content, and save. Useful for seeding initial context before workflows start running.

From workflows - the Memory node in workflows can create and update memories programmatically. This is how memories accumulate over time.

Metadata

Add key/value pairs to any memory: status: active, category: billing, customer_id: 12345.

Workflows can query memories by metadata using exact match filters. All filters are AND-combined - a query for status: active AND category: billing returns only memories matching both.

Use consistent naming. A field called priority in one workflow and importance in another cannot be queried together.

Searching Memories

Two distinct search methods are available, and they serve different purposes:

Metadata query - filters memories by exact field values. Fast and precise. Use when you know the specific value you're looking for.

"Find all memories where customer_id = 12345 and status = active"

Vector search - finds memories by semantic meaning. Returns memories that are conceptually related to your query, even if they use different words. Use when you want to surface relevant context without knowing the exact keywords.

"Find memories related to billing disputes" - returns memories about wrong charges, refund requests, payment failures, even if none use the phrase "billing dispute"

Vector search returns individual chunks (extracted facts), not full documents, with a similarity score from 0 to 1. A threshold of 0.7 is the default - higher values return only closer matches.

Vector search requires the folder to have Generate Embeddings enabled. After enabling, new memories are indexed automatically. Existing memories are indexed in the background.

Propositions and Embedding

When vector search is enabled, Aisle processes each memory through an extraction step before indexing. Rather than embedding the full document text, it extracts discrete facts - called propositions - and indexes each one separately.

This produces more accurate search results. A 10-page document has many distinct facts; embedding the whole document blurs them together. Embedding individual facts surfaces the right one.

For example, a memory containing a customer call summary might produce propositions like:

• "Customer reported duplicate charge on March 15"
• "Customer has been using the product for 18 months"
• "Support agent escalated to billing team"

Each is independently searchable.

Embedding happens in the background after a memory is created or updated. A status indicator shows whether indexing is pending, in progress, completed, or failed.

File Attachments

Memories can include attached files (PDF, Word, Excel, PowerPoint, CSV, and others). When a file is attached, Aisle extracts the text content and uses it as the memory's content for embedding and search. The original file is preserved.

Version History

Every change to a memory is tracked. You can view previous versions, see what changed and when, and restore an earlier version if a workflow writes bad data.

Sharing

By default, memories are private to the folder owner.

To share a folder, click the Sharing button:

• Share with the entire organization or specific team members
• Grant View (read-only) or Edit (full access) permission

Share at the folder level. Individual memory sharing is not available.

Memories as a Knowledge Base in Projects

Memory folders can be assigned to Projects as a knowledge base. When assigned as Reference, the model automatically searches the folder for relevant context during chats. When assigned as Output, workflows running inside the project can save results to the folder.

A folder can serve as both - accumulating findings over time while also surfacing them as context.

Memories as RAG vs Memories via MCP

These are two distinct use cases - both use the same underlying folders but serve different purposes.

Memories as RAG - a memory folder stores documents that a workflow retrieves and injects as context into a prompt. The workflow does the searching; the AI uses the results to give grounded answers. This is internal to Aisle.

Memories via MCP - a memory folder is optionally exposed as an external MCP server. External tools (Claude Code, Cursor, custom agents) connect to it and can read, write, and search the folder directly from outside Aisle. This requires the Enable External MCP Servers toggle to be on in the folder's Settings.

These are independent. A folder can be a RAG source for workflows, an MCP server for external agents, both, or neither.

See Memories MCP Server for setup and available tools.