Framework Adapters

Framework Adapters

SuperLocalMemory V4 ships nine framework adapters. Each adapter implements the memory or history interface that its target framework defines, writing data through the SLM V4 ingestion contract so all records are visible across the CLI, MCP, and dashboard surfaces.

All adapters write to the configured SLM data root. Optional SLM providers, connectors, backup, and downloads retain their separately documented network behavior.

Adapter summary

Framework	PyPI package	Interface	Minimum framework version
LangGraph	langgraph-superlocalmemory	BaseStore (long-term memory)	langgraph >= 1.0.0
Semantic Kernel	semantic-kernel-superlocalmemory	VectorStoreCollection	semantic-kernel >= 1.34.0
Microsoft Agent Framework	agent-framework-superlocalmemory	ContextProvider + HistoryProvider	agent-framework-core >= 1.5.0
LangChain	langchain-superlocalmemory	BaseChatMessageHistory	langchain-core >= 1.0.0
LlamaIndex	llama-index-storage-chat-store-superlocalmemory	BaseChatStore	Python 3.11+
CrewAI	crewai-superlocalmemory	StorageBackend	crewai >= 1.14.6
AutoGen	autogen-superlocalmemory	Memory	autogen-agentchat >= 0.7.5
Google ADK	google-adk-superlocalmemory	BaseMemoryService	google-adk >= 2.5.0
OpenAI Agents	openai-agents-superlocalmemory	SessionABC	openai-agents >= 0.18.3

All adapters require Python 3.11+ and SuperLocalMemory V4 installed in the same virtual environment.

Installation

Install the adapter alongside the target framework:

pip install superlocalmemory <adapter-package>

For example, to add LangGraph support:

pip install superlocalmemory langgraph-superlocalmemory

Source installs are available under ide/integrations/<framework>/ in the repository.

Per-adapter quick reference

LangGraph

Implements BaseStore (long-term memory), providing namespaced key-value storage accessible to create_react_agent and StateGraph.

from langgraph_superlocalmemory import SuperLocalMemoryStore

store = SuperLocalMemoryStore()
store.put(("users", "1"), "profile", {"name": "Ada", "role": "engineer"})
item = store.get(("users", "1"), "profile")
results = store.search(("users",), filter={"role": "engineer"})

Namespace prefix matching is element-wise: ("users",) matches ("users", "1") but not ("users2",). TTL is not supported. Vector similarity search delegates to SLM's native recall.

Semantic Kernel

Implements VectorStoreCollection (SK 1.34+ preview API) for record persistence and CRUD lifecycle operations.

from semantic_kernel_superlocalmemory import SuperLocalMemoryVectorStore

store = SuperLocalMemoryVectorStore()
collection = store.get_collection(Doc, collection_name="docs")
await collection.ensure_collection_exists()
await collection.upsert(Doc(id="d1", text="hello world"))

Search performs field-filter retrieval; dense-vector ANN ranking is a documented follow-up. Validate in CI against the installed SK version — the vector-store API is still marked preview.

Microsoft Agent Framework

Provides SuperLocalMemoryContextProvider and SuperLocalMemoryHistoryProvider for the Agent Framework before_run / after_run hook model (GA, 1.5.0+).

from agent_framework_superlocalmemory import (
    SuperLocalMemoryContextProvider,
    SuperLocalMemoryHistoryProvider,
)

history = SuperLocalMemoryHistoryProvider()
memory = SuperLocalMemoryContextProvider(max_recall=10)
# agent = ChatAgent(..., context_providers=[memory], history_provider=history)

ContextProvider.before_run injects recent session history; wider semantic recall across the SLM store is a documented follow-up.

LangChain

Implements BaseChatMessageHistory. Each session is isolated by a namespaced SHA-256 session identifier; messages are tagged langchain:session:<id>.

from langchain_superlocalmemory import SuperLocalMemoryChatMessageHistory

history = SuperLocalMemoryChatMessageHistory(session_id="my-chat-session")
history.add_messages([HumanMessage(content="What is SLM?"), ...])
for msg in history.messages:
    print(f"{msg.type}: {msg.content}")
history.clear()

All standard LangChain message types are supported. additional_kwargs round-trip through serialization.

LlamaIndex

Implements BaseChatStore. Integrates with ChatMemoryBuffer for automatic conversation management.

from llama_index.storage.chat_store.superlocalmemory import SuperLocalMemoryChatStore
from llama_index.core.memory import ChatMemoryBuffer

chat_store = SuperLocalMemoryChatStore()
memory = ChatMemoryBuffer.from_defaults(
    chat_store=chat_store, chat_store_key="user-123", token_limit=3000
)

Session keys are stored with a llamaindex:<sha256(key)> session identifier.

CrewAI

Implements StorageBackend for scope-aware memory records. CrewAI supplies pre-computed embeddings; the adapter persists them and ranks by cosine similarity in pure Python (no numpy dependency).

from crewai_superlocalmemory import SuperLocalMemoryBackend

backend = SuperLocalMemoryBackend()

Scope-prefix filtering uses hierarchical matching (/project matches /project/alpha, never /projectx). metadata_filter supports equality checks only. For large collections, use SLM's native recall for semantic ranking.

AutoGen

Implements autogen_core.memory.Memory. query() uses SLM's BM25 + semantic recall pipeline, giving the agent access to the full SLM knowledge base, not only AutoGen-added memories.

from autogen_superlocalmemory import SuperLocalMemoryMemory

mem = SuperLocalMemoryMemory()
await mem.add(MemoryContent(content="Ada prefers dark mode.", mime_type=MemoryMimeType.TEXT))
result = await mem.query("Ada preferences")

clear() removes only autogen-mem: rows; personal and shared SLM memories are unaffected.

Note: For Microsoft Agent Framework (agent-framework-core) see the agent_framework_superlocalmemory adapter above; autogen-agentchat and Agent Framework are separate products.

Google ADK

Implements BaseMemoryService. Persists ADK session events and serves semantic recall with post-filtering by app_name + user_id.

from google_adk_superlocalmemory import SuperLocalMemoryService

service = SuperLocalMemoryService()
runner = Runner(agent=agent, app_name="my-app", memory_service=service)

add_session_to_memory is idempotent: re-running for the same session replaces previous events.

OpenAI Agents

Implements SessionABC for ordered conversation history. Items are appended with deterministic ordering (created_at ASC, rowid ASC).

from openai_agents_superlocalmemory import SLMSession
from agents import Runner

session = SLMSession(session_id="user-42-conv-7")
runner = Runner(session=session)
result = await runner.run("What is the capital of France?")

Multiple sessions sharing the same database do not interfere; the prefix used in SQL scans is escaped to prevent wildcard collisions.

Common behaviors

All nine adapters:

• Write through the SLM V4 ingestion contract — records are immediately queryable via slm recall, the dashboard, and any MCP-connected client.
• Isolate sessions with a SHA-256-namespaced or scoped session identifier.
• Handle async via asyncio.to_thread over the synchronous SLM engine, so async agents are never blocked on memory I/O.
• Store data at the configured SLM data root (default ~/.superlocalmemory/memory.db). Pass db_path= to override.

Known limitations across adapters

Limitation	Applies to
ANN vector search is not indexed; search uses SLM recall or in-process cosine	LangGraph, Semantic Kernel, LangChain, LlamaIndex, CrewAI, AutoGen
TTL not supported	LangGraph
metadata_filter equality-only	CrewAI
SK vector-store API is still marked preview	Semantic Kernel
score not exposed in SearchMemoryResponse	Google ADK

---

Part of Qualixar | Created by Varun Pratap Bhardwaj