Cannot provision memory store in Azure AI Foundry for agents despite all infra in place

Question

We’re trying to provision an Azure AI Foundry Agent memory store, but creation fails consistently.

We tested two creation paths:

1. Python SDK with AIProjectClient and client.beta.memory_stores.create(...)
2. Direct REST call to POST /api/projects/{project}/memory_stores?api-version=2025-11-15-preview using AAD token for https://ai.azure.com/

Both paths return the same backend failure: HTTP 500 server_error during memory store creation.

Current environment/config:

• Region alignment: Foundry account, Foundry project, Cosmos DB (NoSQL), Azure AI Search, and Storage are all in East US
• Project capability host: present, provisioning state Succeeded
• Cosmos connection auth type: Microsoft Entra ID
• Cosmos throughput mode for memory containers: fixed manual throughput, 1000 RU/s per container
• Azure AI Search: supported paid tier (not free)
• Blob storage connection: configured
• Permissions: project managed identity has Cosmos data-plane read/write, Search data/index permissions, and Blob Data Contributor on storage
• Models used in tests: gpt-4.1 (chat), text-embedding-3-small (embeddings)

Additional signal that connectivity/auth is likely working: Foundry already created the enterprise_memory database and memory-related containers in Cosmos, so the project-to-Cosmos link appears functional.

Given this setup, we expected memory store provisioning to succeed, but it still fails with server_error from both SDK and REST.

Is there any known service-side condition or additional required setting that can cause this failure pattern?

Answer 1

Hi Samuel Robel

I wanted to add the following clarification.

---

Memory Store Provisioning – Key Considerations

Memory Store provisioning is not only a runtime data‑plane operation. It also performs creation and configuration of Cosmos DB containers and Azure AI Search indexes, which requires:

• Control‑plane permissions
• Network reachability from the Foundry capability host

As a result, a setup that successfully validates connections can still fail during provisioning if:

• Resources are behind private endpoints or firewalls, or
• Only data‑plane roles are assigned.

Therefore, 500 – Internal Server Error responses may originate from the following causes, in addition to backend or regional issues.

---

1. Network (VNET) Perspective

• The outbound network path from the Foundry capability host may be blocked.
  Please review:
  • NSG outbound rules
  • Firewall / proxy logs
• Ensure DNS resolution works for all connected resources.
• Whitelist the Azure AI Foundry Project in the firewall of dependent resources (trusted services / instances).
• Consider using a Managed VNET to reduce networking effort and avoid misconfiguration.

Reference:
https://github.com/strudel0209/ai-foundry-infra-networking/blob/main/README.md

---

2. Permissions Perspective

Provisioning requires control‑plane permissions that are not included in “data‑plane read/write” roles.

Please ensure appropriate roles are assigned, such as:

• Azure AI Account Owner
• Azure AI Account Administrator

Reference:
https://learn.microsoft.com/en-us/azure/foundry/concepts/rbac-foundry

---

3. Code Syntax Perspective (Difference Observed)

✅ Working Code (From Support)

options = MemoryStoreDefaultOptions(
    chat_summary_enabled=True,
    user_profile_enabled=True,
    user_profile_details=(
        "Avoid irrelevant or sensitive data, such as age, financials, "
        "precise location, and credentials"
    )
)

# Create memory store
chat_model = MEMORY_STORE_CHAT_MODEL_DEPLOYMENT_NAME
embedding_model = MEMORY_STORE_EMBEDDING_MODEL_DEPLOYMENT_NAME

definition = MemoryStoreDefaultDefinition(
    chat_model=chat_model,
    embedding_model=embedding_model,
    options=options
)

memory_store = project_client.beta.memory_stores.create(
    name=memory_store_name,
    definition=definition,
    description="Memory store for customer support agent",
)

---

⚠️ Code Used by Customer

client = AIProjectClient(
    endpoint=PROJECT_ENDPOINT,
    credential=DefaultAzureCredential(),
    allow_preview=True,
)

memory_store = client.beta.memory_stores.create(
    name=MEMORY_STORE_NAME,
    definition=MemoryStoreDefaultDefinition(
        kind=MemoryStoreKind.DEFAULT,
        chat_model=CHAT_MODEL,
        embedding_model=EMBEDDING_MODEL,
    ),
)

---

4. Resource Provider Perspective

Please ensure the following Azure resource providers are registered in the subscription:

• Microsoft.KeyVault
• Microsoft.CognitiveServices
• Microsoft.Storage
• Microsoft.MachineLearningServices
• Microsoft.Search
• Microsoft.Network
• Microsoft.App
• Microsoft.ContainerService
• (Optional – for Bing Search tool) Microsoft.Bing

5. Regions Perspective

Please test the models in supported regions

Reference:
https://learn.microsoft.com/en-us/azure/foundry/agents/how-to/virtual-networks

<https://learn.microsoft.com/en-us/azure/foundry/reference/region-support#foundry-projects>

I hope this helps in addressing the issue.

Looking forward to hearing from you.

Thank you.

---