v1.2.0

Added

• AgentFrameworkOptions.ExposeMemoryToolsFromContextProvider — optional memory-tool exposure via AIContext.Tools (#86). Neo4jMemoryContextProvider can now surface the six standard memory tools (MemoryToolFactory.CreateAIFunctions()) itself through AIContext.Tools, so AIContextProviders = [memoryProvider] alone is enough to give an agent LLM-callable memory tools — no separate ChatOptions.Tools = [.. memoryTools] wiring required.
  Defaults to false: AddAgentMemoryFramework registers MemoryToolFactory unconditionally, and its tools include write-capable ones (remember_fact, remember_preference), so exposure must stay opt-in rather than firing just because the factory is present in DI. Every BuildContextAsync branch (a recall hit, an empty recall, a recall failure, or no user message at all) shares one AIContext-construction helper, so tool availability never silently drops on a quiet turn.
• ContextFormatOptions.MaxChatHistoryMessages (#91, see Fixed below) — the new name for what was MaxContextMessages.

Fixed

• AddGraphRagAdapter() could never actually resolve IGraphRagContextSource — IDriver was never registered in DI. Neo4jGraphRagContextSource takes IDriver via constructor injection, but AddNeo4jAgentMemory only registered INeo4jDriverFactory (which wraps IDriver via GetDriver()), never IDriver itself. Any host that called AddGraphRagAdapter() hit InvalidOperationException: Unable to resolve service for type 'Neo4j.Driver.IDriver' the first time GraphRAG retrieval ran — this broke the feature for every consumer, not just the sample that surfaced it. Fixed by registering IDriver from the existing INeo4jDriverFactory singleton in AddNeo4jAgentMemory.
• ContextFormatOptions.MaxContextMessages renamed to MaxChatHistoryMessages, old name kept as an [Obsolete] compatibility alias (#91). The option was documented as capping the complete injected context, but the implementation always preserved the context prefix and every memory-derived block (entities/facts/preferences/reasoning traces/GraphRAG) — only recalled chat history was ever truncated to fit. The renamed option's implementation now matches its name: it bounds only recalled chat history, independent of the prefix/memory-block count. Negative values are rejected by option validation; MaxChatHistoryMessages = 0 means no recalled chat history, but memory blocks may still be included. Use ContextBudget.MaxTokens/MaxCharacters for a hard cap on total prompt size.

Changed

• Every sample now calls a real Azure OpenAI chat and/or embedding model — no mocks. Removed EchoChatClient/ShoppingEchoChatClient/StubEmbeddingGenerator from every sample (AgentWithMemory, RealAgent, MemoryToolsAgent, ChatHistoryProvider, ShoppingAssistant, BlendedAgent, MinimalAgent, McpHost, and the AspireDemo app). A new samples/AgentMemory.Samples.Shared project centralizes the wiring:

  • RealAzureOpenAI.TryCreate resolves AZURE_OPENAI_* environment variables and fails fast with setup instructions if they're missing (no silent mock fallback).
  • MemoryTraceChatClient prints the <recalled_memory> context the provider injects before each live model call.
  • SampleConsole gives each sample color-coded user/assistant/memory-action console output.

  The five agent samples now let the model decide on its own when to call memory/product tools, rather than scripting the calls a mock model couldn't make itself.

Full Changelog: see CHANGELOG.md