10xGraphv0.9.2Search

    Memory Tools

    In shortmemory_tool, user_memory_tool, and agent_memory_tool give an agent long-term memory to store, search, update, and delete facts.

    • 3 min read
    • 4 sections
    • Updated
    • v0.9.2
    • Markdown

    Prebuilt tools that give an agent access to long-term memory — the ability to store, search, update, and delete facts across conversations.

    Import path: tenxgraph.prebuilt.tools

    There are three tools, each for a different memory integration path:

    Tool Path Operations
    memory_tool Manual / MemoryIntegration wiring search, store, update, delete
    user_memory_tool Agent(memory=MemoryConfig(...)) search, remember
    agent_memory_tool Agent(memory=MemoryConfig(...)) search (read-only)

    All three tools require a configured BaseStore (e.g. a Qdrant-backed store) injected through the DI container or passed explicitly.

    memory_tool

    The general-purpose LLM-callable memory tool. Use this when wiring memory manually or through MemoryIntegration.

    Operations

    action Required fields Description
    search query Semantic search across all memories for the current user
    store content, memory_key Save a new memory (auto-updates if memory_key already exists)
    update memory_id, content Overwrite a specific memory by ID
    delete memory_id Remove a specific memory by ID

    Parameters

    Parameter Type Default Description
    action str "search" One of search, store, update, delete
    content str "" Text to store or update
    memory_key str "" Short snake_case key used for dedup (e.g. "user_name")
    memory_id str "" ID of the memory to update or delete
    query str "" Search query
    memory_type str None Memory type ("episodic", "semantic", etc.)
    category str None Category label for filtering
    limit int 5 Maximum number of search results
    score_threshold float None Minimum similarity score for search
    write_mode str "merge" "merge" or "replace" on update

    Notes

    • Write operations (store, update, delete) are scheduled as background tasks and return {"status": "scheduled"} immediately.
    • Search flushes pending writes first so results are always up to date.
    • The memory_key field enables automatic deduplication: if a memory with the same key exists it is updated rather than duplicated.

    Usage

    Python
    from tenxgraph.prebuilt.tools.memory import memory_tool
from tenxgraph.core.graph import Agent, ToolNode
from tenxgraph.storage import QdrantStore  # or any BaseStore subclass

store = QdrantStore(...)

agent = Agent(
    model="gpt-4o-mini",
    tool_node=ToolNode([memory_tool]),
    system_prompt=[{
        "role": "system",
        "content": (
            "You have long-term memory. "
            "Always search memory at the start of a conversation. "
            "Store important facts about the user after each interaction."
        ),
    }],
)
app = agent.compile(store=store)

    user_memory_tool (factory)

    Created via make_user_memory_tool(memory_config). Used automatically by Agent(memory=MemoryConfig(...)) — you do not normally need to instantiate it yourself.

    Operations

    action Required fields Description
    search text Semantic search over user-scoped memories
    remember text Save a user fact or preference

    Parameters

    Parameter Type Default Description
    action str "search" "search" or "remember"
    text str required Query text or text to remember
    memory_type str None Override the configured memory type
    category str None Override the configured category
    limit int None Override the configured result limit

    Usage via MemoryConfig

    Python
    from tenxgraph.core.graph import Agent
from tenxgraph.storage.store.memory_config import MemoryConfig, UserMemoryConfig

agent = Agent(
    model="gpt-4o-mini",
    memory=MemoryConfig(
        store=store,
        user_memory=UserMemoryConfig(enabled=True),
    ),
)
# The user_memory_tool is registered automatically.
app = agent.compile(store=store)

    agent_memory_tool (factory)

    Created via make_agent_memory_tool(memory_config). Read-only — the LLM can search agent-scoped or app-scoped memories but cannot write them.

    Operations

    Parameter Required Description
    query required Semantic search query
    memory_type None Override configured memory type
    category None Override configured category
    limit None Override result limit

    Usage via MemoryConfig

    Python
    from tenxgraph.storage.store.memory_config import MemoryConfig, AgentMemoryConfig

agent = Agent(
    model="gpt-4o-mini",
    memory=MemoryConfig(
        store=store,
        agent_memory=AgentMemoryConfig(
            enabled=True,
            agent_id="my-agent-v1",
        ),
    ),
)
app = agent.compile(store=store)

    Example: manual wiring with memory_tool

    Python
    from tenxgraph.prebuilt.tools.memory import memory_tool
from tenxgraph.prebuilt.agent import ReactAgent
from tenxgraph.storage import create_local_qdrant_store
from tenxgraph.storage.store.embedding import OpenAIEmbedding

store = create_local_qdrant_store(
    path="./memory_db",
    embedding=OpenAIEmbedding(model="text-embedding-3-small"),
)

agent = ReactAgent(
    model="gpt-4o-mini",
    tools=[memory_tool],
    system_prompt=[{
        "role": "system",
        "content": (
            "You have persistent memory. At the start of every conversation, "
            "call memory_tool with action='search' to recall relevant context. "
            "After the conversation, store important new facts with action='store'."
        ),
    }],
)
app = agent.compile(store=store)

result = await app.ainvoke(
    {"message": "My name is Alice and I prefer Python."},
    config={"thread_id": "t1", "user_id": "alice"},
)
    Last updated for v0.9.2Edit this page on GitHubReport an issue