Program.cs

// Copyright (c) Microsoft. All rights reserved.

// This sample shows how to add a basic custom memory component to an agent.
// The memory component subscribes to all messages added to the conversation and
// extracts the user's name and age if provided.
// The component adds a prompt to ask for this information if it is not already known
// and provides it to the model before each invocation if known.

using System.Text;
using System.Text.Json;
using Azure.AI.Projects;
using Azure.Identity;
using Microsoft.Agents.AI;
using Microsoft.Extensions.AI;
using SampleApp;

var endpoint = Environment.GetEnvironmentVariable("FOUNDRY_PROJECT_ENDPOINT") ?? throw new InvalidOperationException("FOUNDRY_PROJECT_ENDPOINT is not set.");
var model = Environment.GetEnvironmentVariable("FOUNDRY_MODEL") ?? "gpt-5.4-mini";

// WARNING: DefaultAzureCredential is convenient for development but requires careful consideration in production.
// In production, consider using a specific credential (e.g., ManagedIdentityCredential) to avoid
// latency issues, unintended credential probing, and potential security risks from fallback mechanisms.
var projectClient = new AIProjectClient(new Uri(endpoint), new DefaultAzureCredential());

// Create a separate IChatClient for the memory component to use for structured extraction.
// The memory component calls the model with a ResponseFormat (JSON schema) to extract user info.
// Using a dedicated client here avoids mixing side-channel extraction calls with the agent's
// conversation history, and avoids the chicken-and-egg problem of needing an IChatClient
// before the main agent is constructed.
IChatClient extractionClient =
    new AIProjectClient(
        new Uri(endpoint),
        new DefaultAzureCredential())
    .GetProjectOpenAIClient()
    .GetResponsesClient()
    .AsIChatClient(model);

// Create the agent with instructions and the custom memory context provider.
// The memory component is attached to all sessions created by the agent. Here each new memory
// component will have its own user info object, so each session will have its own memory.
// In real world applications/services, where the user info would be persisted in a database,
// and preferably shared between multiple sessions used by the same user, ensure that the
// factory reads the user id from the current context and scopes the memory component
// and its storage to that user id.
AIAgent agent = projectClient.AsAIAgent(new ChatClientAgentOptions
{
    ChatOptions = new ChatOptions
    {
        ModelId = model,
        Instructions = "You are a friendly assistant. Always address the user by their name.",
    },
    AIContextProviders = [new UserInfoMemory(extractionClient)]
});

// Create a new session for the conversation.
AgentSession session = await agent.CreateSessionAsync();

Console.WriteLine(">> Use session with blank memory\n");

// Invoke the agent and output the text result.
Console.WriteLine(await agent.RunAsync("Hello, what is the square root of 9?", session));
Console.WriteLine(await agent.RunAsync("My name is Ruaidhrí", session));
Console.WriteLine(await agent.RunAsync("I am 20 years old", session));

// We can serialize the session. The serialized state will include the state of the memory component.
JsonElement sessionElement = await agent.SerializeSessionAsync(session);

Console.WriteLine("\n>> Use deserialized session with previously created memories\n");

// Later we can deserialize the session and continue the conversation with the previous memory component state.
var deserializedSession = await agent.DeserializeSessionAsync(sessionElement);
Console.WriteLine(await agent.RunAsync("What is my name and age?", deserializedSession));

Console.WriteLine("\n>> Read memories using memory component\n");

// It's possible to access the memory component via the agent's GetService method.
var userInfo = agent.GetService<UserInfoMemory>()?.GetUserInfo(deserializedSession);

// Output the user info that was captured by the memory component.
Console.WriteLine($"MEMORY - User Name: {userInfo?.UserName}");
Console.WriteLine($"MEMORY - User Age: {userInfo?.UserAge}");

Console.WriteLine("\n>> Use new session with previously created memories\n");

// It is also possible to set the memories using a memory component on an individual session.
// This is useful if we want to start a new session, but have it share the same memories as a previous session.
var newSession = await agent.CreateSessionAsync();
if (userInfo is not null && agent.GetService<UserInfoMemory>() is UserInfoMemory newSessionMemory)
{
    newSessionMemory.SetUserInfo(newSession, userInfo);
}

// Invoke the agent and output the text result.
// This time the agent should remember the user's name and use it in the response.
Console.WriteLine(await agent.RunAsync("What is my name and age?", newSession));

namespace SampleApp
{
    /// <summary>
    /// Sample memory component that can remember a user's name and age.
    /// </summary>
    internal sealed class UserInfoMemory : AIContextProvider
    {
        private readonly ProviderSessionState<UserInfo> _sessionState;
        private IReadOnlyList<string>? _stateKeys;
        private readonly IChatClient _chatClient;

        public UserInfoMemory(IChatClient chatClient, Func<AgentSession?, UserInfo>? stateInitializer = null)
        {
            this._sessionState = new ProviderSessionState<UserInfo>(
                stateInitializer ?? (_ => new UserInfo()),
                this.GetType().Name);
            this._chatClient = chatClient;
        }

        public override IReadOnlyList<string> StateKeys => this._stateKeys ??= [this._sessionState.StateKey];

        public UserInfo GetUserInfo(AgentSession session)
            => this._sessionState.GetOrInitializeState(session);

        public void SetUserInfo(AgentSession session, UserInfo userInfo)
            => this._sessionState.SaveState(session, userInfo);

        protected override async ValueTask StoreAIContextAsync(InvokedContext context, CancellationToken cancellationToken = default)
        {
            var userInfo = this._sessionState.GetOrInitializeState(context.Session);

            // Try and extract the user name and age from the message if we don't have it already and it's a user message.
            if ((userInfo.UserName is null || userInfo.UserAge is null) && context.RequestMessages.Any(x => x.Role == ChatRole.User))
            {
                // The Foundry Responses API requires the model name in the request body.
                // Retrieve it from the client's metadata so callers don't need to pass it separately.
                var modelId = this._chatClient.GetService<ChatClientMetadata>()?.DefaultModelId
                    ?? throw new InvalidOperationException(
                        "Could not retrieve DefaultModelId from the extraction IChatClient. " +
                        "Ensure the client was created with a model ID (e.g., via projectClient.AsAIAgent(...)).");
                var result = await this._chatClient.GetResponseAsync<UserInfo>(
                    context.RequestMessages,
                    new ChatOptions()
                    {
                        ModelId = modelId,
                        Instructions = "Extract the user's name and age from the message if present. If not present return nulls."
                    },
                    cancellationToken: cancellationToken);

                userInfo.UserName ??= result.Result.UserName;
                userInfo.UserAge ??= result.Result.UserAge;
            }

            this._sessionState.SaveState(context.Session, userInfo);
        }

        protected override ValueTask<AIContext> ProvideAIContextAsync(InvokingContext context, CancellationToken cancellationToken = default)
        {
            var userInfo = this._sessionState.GetOrInitializeState(context.Session);

            StringBuilder instructions = new();

            // If we don't already know the user's name and age, add instructions to ask for them, otherwise just provide what we have to the context.
            instructions
                .AppendLine(
                    userInfo.UserName is null ?
                        "Ask the user for their name and politely decline to answer any questions until they provide it." :
                        $"The user's name is {userInfo.UserName}.")
                .AppendLine(
                    userInfo.UserAge is null ?
                        "Ask the user for their age and politely decline to answer any questions until they provide it." :
                        $"The user's age is {userInfo.UserAge}.");

            return new ValueTask<AIContext>(new AIContext
            {
                Instructions = instructions.ToString()
            });
        }
    }

    internal sealed class UserInfo
    {
        public string? UserName { get; set; }
        public int? UserAge { get; set; }
    }
}