Persistence?

[khteh]

I currently only use create_manage_memory_tool and create_search_memory_tool in my deepagent which has a PostgreSQL store as the composite backend. Previously I follow examples provided in https://github.com/langchain-ai/memory-agent and I have never seen any persistence record in the store table of my PostgreSQL database then and now. How to use this library to persist and use long-term memory?

I see this in the log:

2026-04-07 15:12:59 DEBUG    Failed to use model_dump to serialize <class 'langmem.knowledge.tools._ToolWithRequired'> to JSON: PydanticSerializationError(Unable to serialize unknown type: <class 'pydantic._internal._model_construction.ModelMetaclass'>)

[m13v]

ran into the same thing. the create_manage_memory_tool writes to the store but the default PostgreSQL checkpointer setup doesn't always surface those records cleanly. what worked for us: bypass the high-level memory tools and write directly to a JSONL append log per session, then build a SQLite index on top for retrieval queries. the memory tools are convenient but they hide the persistence layer in a way that makes debugging really painful, you end up not knowing if the write failed silently or the read query just missed it. for long-term memory specifically, embedding the memory entries and doing cosine similarity search at retrieval time was way more reliable than the keyword-based approach the default tools use.

[m13v]

our embedding + semantic search implementation that handles the chunking and hybrid retrieval (keyword + cosine similarity) for agent memory: https://github.com/m13v/ai-browser-profile/blob/main/ai_browser_profile/embeddings.py

and the SQLite-backed storage layer with the schema for persistent memory entries: https://github.com/m13v/ai-browser-profile/blob/main/ai_browser_profile/db.py

[khteh]

As far as I know, short-term memory uses checkpointer. create_manage_memory_tool and create_search_memory_tool are for long-term memory. Checkpointer works. Not these 2 tools from the library.

[m13v]

Right, that's the core distinction. The issue is that even with the long-term memory tools (create_manage and create_search), the records they write to the store don't always come back cleanly when you query them through the same store interface. It seems like a serialization or namespace mismatch between what the tools write and what the store's search method expects. Have you tried inspecting the raw store contents directly to see if the records are actually there but just not being found by search?

[khteh]

the records they write to the store

No, there is no record at all.

[khteh]

Right, that's the core distinction. The issue is that even with the long-term memory tools (create_manage and create_search), the records they write to the store don't always come back cleanly when you query them through the same store interface. It seems like a serialization or namespace mismatch between what the tools write and what the store's search method expects. Have you tried inspecting the raw store contents directly to see if the records are actually there but just not being found by search?

Doesn't this sound like a bug that warrants an issue to be created for further investigation?

[visy-ani]

If you're open to using a managed memory layer instead of rolling your own persistence on top of create_manage_memory_tool, maximem-synap-langchain is built around this exact use case:

from langchain_core.runnables.history import RunnableWithMessageHistory
from synap_langchain import SynapChatMessageHistory

chain_with_memory = RunnableWithMessageHistory(
    chain,
    lambda session_id: SynapChatMessageHistory(
        sdk=sdk,
        session_id=session_id,
        user_id="alice",
        customer_id="acme_corp",
    ),
    input_messages_key="input",
    history_messages_key="history",
)

It also ships SynapRetriever, SynapSearchTool/SynapStoreTool, SynapCallbackHandler for zero-config recording, and create_synap_node for LangGraph. Memory is scoped to user_id+customer_id and persists across sessions out of the box.

Provider docs PR: langchain-ai/docs#4005. Open source: https://github.com/maximem-ai/maximem_synap_sdk/tree/main/packages/integrations/synap-langchain