Skip to main content

Integrations hub: Chat SDK, CrewAI, MCP tools, and retrieval upgrades

· 2 min read
Anurag Atulya

Integration and retrieval improvements: give chatbots memory, wire CrewAI to Atulya banks, tune MCP exposure, and trace observations back to facts.

Model leaderboard​

Public Model Leaderboard ranks supported LLMs on retain and reflect for accuracy, speed, and cost.

Chat SDK integration​

Persistent memory for AI SDK useChat:

import { useChat } from "ai/react";
import { createAtulyaMiddleware } from "atulya-sdk/chat";

const { messages, input, handleSubmit } = useChat({
  middleware: createAtulyaMiddleware({
    bankId: "user-123",
    serverUrl: "http://localhost:8888",
  }),
});

Middleware retains turns and injects recalled context. Chat SDK docs

CrewAI integration​

from crewai import Crew
from atulya_integrations.crewai import AtulyaMemory

crew = Crew(
    agents=[...],
    tasks=[...],
    memory=True,
    memory_config={
        "provider": "atulya",
        "config": {"bank_id": "my-crew", "server_url": "http://localhost:8888"},
    },
)

Full retrieval stack (semantic, BM25, graph, rerank), not in-process SQLite. CrewAI docs

Configurable MCP tools​

Restrict exposed tools per deployment:

export ATULYA_API_MCP_ENABLED_TOOLS="recall,retain"

Expanded MCP surface with richer parameters on existing tools.

Batch observations consolidation​

Multiple observations consolidated in one batched LLM pass after heavy ingests (~10x faster on consolidation-heavy workloads). Automatic, no config.

ZeroEntropy reranker​

export ATULYA_API_RERANKER_PROVIDER=zeroentropy
export ATULYA_API_ZEROENTROPY_API_KEY=your-key

Observation source facts​

Recalled observations include underlying source facts:

results = client.recall(bank_id="my-bank", query="user preferences")
for result in results:
    if result.type == "observation":
        print(result.text)
        for fact in result.source_facts:
            print(" -", fact.text)

Reliability and performance notes​

  • Bank config API enabled by default
  • OpenClaw: autoRecall toggle, provider exclusions, auth-aware health checks
  • Richer reflect/retain/consolidation configuration
  • Source document metadata on extracted facts
  • Clear error when embedding dimensions exceed pgvector HNSW limits
  • Multi-tenant schema isolation fixes
  • Lower recall memory footprint
  • Default OpenAI LLM: gpt-4o-mini for new deployments
  • MCP spec alignment; Docker named volume startup fix
  • Reranker graceful degradation on upstream errors
  • Sharper temporal ordering for facts
  • Documents tracked even when extraction returns zero facts

Changelog