AG2 Rewrites Group Chat, Mem0 Goes Graph-Backed, RAGAS Hits 1.0

<p><strong>AG2 0.6</strong> ships a replayable group-chat API that finally makes multi-agent regression testing practical, while <strong>Mem0 v2</strong> upgrades agent memory from flat embeddings to a property graph that tracks entity relationships over time. <strong>RAGAS 1.0</strong> locks in a stable RAG evaluation API teams can safely pin in CI — wel...

Highlights

  • AG2 0.6 ships a ground-up rewrite of group-chat with typed handoffs and a replay() API that re-runs prior conversations against updated agents — making multi-agent regression testing tractable for the first time (AG2)
  • Mem0 v2 replaces its flat key-value memory store with a pluggable property-graph layer, letting agents query entity relationships and temporal chains of prior facts rather than isolated semantic snippets (Mem0)
  • RAGAS 1.0 reaches a stable API milestone, cementing the leading open-source RAG evaluation library as a pinnable production dependency with frozen scorer interfaces (RAGAS)
  • LangGraph Platform ships per-run cost tracking alongside its recently launched MCP server support, giving operators token-spend visibility per graph execution without custom instrumentation (LangGraph)

Key Signals

  1. AG2 0.6 makes multi-agent conversations testable and replayable — June 2026
    AutoGen's original group-chat stored conversation state in mutable Python objects, making deterministic test harnesses nearly impossible to write. AG2 0.6 introduces ConversationState with typed HandoffMessage objects and a replay() call that re-executes stored runs against modified agent configs — unlocking regression pipelines for long-horizon workflows. Teams on AutoGen-era code need to migrate handoff logic; AG2 ships a compatibility shim for existing GroupChatMessage dict-based code. (AG2)
  2. Mem0 v2 brings graph-backed long-term memory to agents — June 2026
    Mem0 v1 stored facts as flat embeddings, limiting recall to semantic similarity over isolated snippets. Version 2 introduces a property graph layer tracking entities, relationships, and temporal chains — so an agent can query "what did the user say about project X after meeting Y" instead of just "anything about project X." The graph backend is pluggable: Neo4j and Kuzu for production, an in-process SQLite graph for local development. (Mem0)
  3. RAGAS 1.0 marks RAG evaluation's production-readiness moment — June 2026
    After eighteen months of pre-1.0 iteration, RAGAS tagged a stable release with a frozen public API, async evaluation support, and official integration hooks for LlamaIndex, LangChain, and Haystack. The milestone matters operationally: teams can now pin ragas==1.0.* in CI without worrying about breaking metric renames or scorer interface churn that plagued earlier versions — and the timing dovetails with OpenAI Evals sunsetting this fall. (RAGAS)

Why It Matters / What To Watch

  1. Stateful, testable agent memory is within reach for mid-sized teams
    • Mem0 v2's graph backend and AG2's replay API together address the two hardest parts of production agent reliability: "does my agent recall context correctly across sessions?" and "did my code change break long-horizon behavior?" (Mem0) (AG2)
    • Watch for LlamaIndex and LangGraph to surface Mem0 v2 as a first-class memory backend option — both expose plugin APIs designed for exactly this kind of swap.
  2. Eval tooling is stabilizing just as the OpenAI Evals gap widens
    • RAGAS 1.0 and LangGraph's per-run cost tracking give teams a credible eval-plus-observability stack that doesn't depend on third-party SaaS — a direct answer to the OpenAI Evals November 30 shutdown announced June 6. (RAGAS) (LangGraph)
    • Teams building internal RAG benchmarks should migrate away from ad-hoc metric scripts and onto the RAGAS 1.0 scorer interface now; the earlier you pin, the less churn when the OpenAI Evals deadline lands.

Quick Links