Memori 0.3.0

Memori

Durable memory for AI applications built on Microsoft.Extensions.AI.

Memori gives your AI app a persistent, searchable memory layer: capture conversations, recall relevant facts, and inject context into prompts — without coupling your code to a specific database or LLM provider.

Installation

dotnet add package Memori

What You Get

• IConversationStorage for conversation/session/entity/process data — bring your own backend
• VectorStoreCollection<string, MemoryFactRecord> for durable fact storage via any Microsoft.Extensions.VectorData provider
• Built-in InMemoryConversationStorage and InMemoryVectorStore for tests, demos, and local development
• IEmbeddingGenerator<string, Embedding<float>> from Microsoft.Extensions.AI as the embedding surface
• A Memori facade for attribution, session tracking, capture, recall, and augmentation in one place
• IChatClient middleware that wires recall and capture into any Microsoft.Extensions.AI-compatible provider
• Scope isolation for multi-tenant memory
• Versioning and conflict resolution (last-write-wins, merge, manual)
• Thread summarization via ChatClientThreadSummarizer
• Memory management APIs for user-facing inspection, search, edit, soft-delete, and restore

Basic Example

using Memori.Models;
using Memori.Search;
using Memori.Storage;

var conversationStorage = new InMemoryConversationStorage();
var vectorStore = new InMemoryVectorStore();
var factCollection = vectorStore.GetCollection<string, MemoryFactRecord>("memori_facts");

var entityId = await conversationStorage.GetOrCreateEntityAsync("user_123");

await factCollection.UpsertAsync(new MemoryFactRecord
{
 Id = Guid.NewGuid().ToString("N"),
 EntityId = entityId,
 Content = "The user's favorite color is blue",
 CreatedAt = DateTimeOffset.UtcNow
});

var search = new MemorySearchService(factCollection);

var results = await search.RecallAsync(entityId, "What is my favorite color?");
Console.WriteLine(search.FormatPromptContext(results));

Facade Usage

using Memori;
using Memori.Augmentation;
using Memori.Models;
using Memori.Storage;

var conversationStorage = new InMemoryConversationStorage();
var vectorStore = new InMemoryVectorStore();
var factCollection = vectorStore.GetCollection<string, MemoryFactRecord>("memori_facts");

var memori = new MemoriEngine(
 conversationStorage,
 factCollection,
 new MemoriOptions { StripSystemMessagesOnCapture = true },
 augmentationClient: new NullAugmentationClient());

memori.Attribution("user_123", "support_agent");
memori.SetSession("session_abc");

await memori.CaptureAsync(new[]
{
 new ConversationMessage(ConversationRoles.User, "My favorite color is blue."),
 new ConversationMessage(ConversationRoles.Assistant, "Noted.")
});

var recalled = await memori.RecallAsync("What is my favorite color?");
await memori.WaitForAugmentationAsync();

var promptContext = memori.BuildPromptContext(recalled);
Console.WriteLine(promptContext.RenderedText);

Scope Isolation

memori.Attribution("user_123");
memori.SetScope("workspace-a");
var results = await memori.RecallAsync("coffee");
memori.ClearScope();

Microsoft.Extensions.AI Middleware

using Memori;
using Microsoft.Extensions.AI;

IChatClient innerClient = /* your provider-backed IChatClient */;

IChatClient client = new ChatClientBuilder(innerClient)
 .UseMemori(memori)
 .Build(serviceProvider);

Recalled memory is injected as a system message before the existing chat history by default. Injection placement and role are configurable.

Dependency Injection

services.AddMemori(options =>
{
 options.SessionTimeout = TimeSpan.FromMinutes(30);
 options.PromptInjectionPlacement = PromptInjectionPlacement.AfterSystemAndDeveloperMessages;
});

var memori = serviceProvider.CreateMemori();

Bind from configuration:

services.AddMemori(configuration.GetSection("Memori"));

Supply custom conversation storage, embedding, and augmentation through factories:

services.AddMemori(
 sp => new MyConversationStorage(),
 configureOptions: options => { options.RecallRelevanceThreshold = 0.2; });

Augmentation

Augmentation extracts structured memory (facts, semantic triples, process attributes, summaries) from captured conversations in the background.

• NullAugmentationClient: no-op, for hosts that only want capture/recall.
• PromptAugmentationClient: uses an IChatClient to extract structured JSON output.
• Implement IAugmentationClient for custom extraction logic.

Versioning

var versioning = new VersioningService(ConflictResolutionStrategy.LastWriteWins);
var resolution = versioning.ResolveConflict(incoming, existing, expectedVersion: 1);

Supports last-write-wins, merge, and manual resolution strategies with audit trail.

Thread Summarization

var summarizer = new ChatClientThreadSummarizer(chatClient);
var summary = await summarizer.SummarizeAsync(messages);
var updated = await summarizer.SummarizeAsync(newMessages, previousSummary);

Memory Management

var management = serviceProvider.GetRequiredService<IMemoryManagementService>();
var memories = await management.ListMemoriesAsync("entity-1");
await management.SoftDeleteMemoryAsync("fact-id");
await management.RestoreMemoryAsync("fact-id");
await management.HardDeleteMemoryAsync("fact-id");

Learn More

• Full feature overview:
• Developer guide:
• Architecture and design notes:

Status

Core primitives, distributed ranking, composite collection, scope isolation, versioning, thread summarization, and memory management are all implemented. 211 NUnit tests. No first-party database integrations — implement IConversationStorage and supply a VectorStore provider in your own package.

License

Apache-2.0