Semantic Memory: AI with a Recall

Sire isn't just an execution engine---it's an AI-native platform that learns from every interaction. At the core of this learning is our Hierarchical Semantic Memory.

How Memory Works

Every time a workflow step finishes, Sire doesn't just store the output in a database; it indexes it.

1. Vectorization: Step outputs (JSON, text, status codes) are converted into mathematical representations called "embeddings."
2. Storage: These embeddings are stored in a high-performance vector store.
3. Recall: When you prompt a new mission, Sire searches this vector store for past results, configurations, and successes that are semantically similar to your current goal.

Hierarchical Scoping

To balance power with privacy, Sire organizes memory into three distinct layers:

1. Assistant Memory (Private)

• Scope: Only results from the current assistant's past executions.
• Value: Great for remembering user-specific preferences or state within a single long-running project.

2. Workflow Memory (Pattern-Based)

• Scope: Results from all executions of a specific workflow.
• Value: If you run a "Market Research" workflow weekly, Sire remembers the sources it found last week and can detect trends or avoid redundant searches.

3. Tenant Memory (Organization-Wide)

• Scope: Results from any execution across your entire organization.
• Value: If a colleague successfully integrated a new MCP tool yesterday, Sire can recall that success to help you set up your own workflow today.

Using Memory Search in Workflows

You can also explicitly use the memory in your DAGs with the built-in sire:local/memory.search tool.

• Prompt-based Retrieval: "Search for the last 5 successful summaries of company reports."
• Context Injection: "Find the API key reference from our previous deployment and pass it to this step."
• Scoping: Specify if you want to search assistant, workflow, or tenant level data.

Privacy & Security

We understand that memory is powerful---and sensitive.

• Isolation: Tenant Memory never crosses organization boundaries. Your data is always isolated to your specific tenant ID.
• Audit Logging: Every memory search is logged, so you can see when and why the AI recalled specific information.
• Opt-Out: You can configure specific workflows or steps to bypass indexing if they handle highly sensitive or temporary data.

The Value of Memory

Faster Execution

By recalling past successes, Sire can skip redundant steps or provide the AI with better starting parameters.

Improved Accuracy

The AI learns from previous failures. If a specific tool call failed three times last week, the generator will attempt a different approach today.

Organizational Intelligence

Sire becomes a living repository of your team's collective knowledge, making automation smarter as you use it.