Add Mem0 integration - support for Mem0 platform #391

Amnah199 · 2025-10-28T14:56:56Z

Related Issues

partially fixes - #219

Proposed Changes:

This PR integrates mem0 as a memory store in Haystack, supporting basic add, retrieve, and delete operations.
It updates the memory_agents.Agent class to enable retrieving and persisting long-term conversational context.

Key Changes

Added memory support to memory_agents.Agent.
Before each execution, relevant memories are fetched from Mem0MemoryStore using the latest user message as the query.
Retrieved memories are injected into the conversation as [MEMORY #n] ChatMessage objects.
When memories exist, the system prompt is updated to tell the LLM that [MEMORY] messages represent long-term context.
After each run, new user and assistant messages are saved to the memory store.
Memory is optional; agents behave normally when memory_store=None.

How did you test it?

Unit and Integration Tests
Tested locally with examples

Notes for the reviewer

Support for Memory.from_config will be revised in a separate PR as it has a slightly different implementation than mem0 MemoryClient.
https://docs.mem0.ai/platform/platform-vs-oss

Checklist

I have read the contributors guidelines and the code of conduct
I have updated the related issue with new insights and changes
I added unit tests and updated the docstrings
I've used one of the conventional commit types for my PR title: fix:, feat:, build:, chore:, ci:, docs:, style:, refactor:, perf:, test:.
I documented my code
I ran pre-commit hooks and fixed any issue

coveralls · 2025-10-28T14:59:24Z

Pull Request Test Coverage Report for Build 20066169239

Details

0 of 0 changed or added relevant lines in 0 files are covered.
94 unchanged lines in 1 file lost coverage.
Overall coverage decreased (-2.2%) to 71.082%

Files with Coverage Reduction	New Missed Lines	%
components/agents/agent.py	94	53.88%

Totals
Change from base Build 20012634360:	-2.2%
Covered Lines:	1288
Relevant Lines:	1812

💛 - Coveralls

…ental into mem0-integration

haystack_experimental/components/memory_agents/agent.py

haystack_experimental/memory_stores/examples/memory_store_with_agent.py

pyproject.toml

sjrl · 2026-01-16T10:29:40Z

haystack_experimental/components/agents/agent.py

        confirmation_strategies: Optional[dict[str, ConfirmationStrategy]] = None,
        tool_invoker_kwargs: Optional[dict[str, Any]] = None,
        chat_message_store: Optional[ChatMessageStore] = None,
+        memory_store: Optional["Mem0MemoryStore"] = None,


This should use the type defined in memory_stores.types.protocol

sjrl · 2026-01-16T10:30:17Z

haystack_experimental/memory_stores/examples/memory_store_with_agent.py

+# SPDX-FileCopyrightText: 2022-present deepset GmbH <[email protected]>
+#
+# SPDX-License-Identifier: Apache-2.0


I think we should move this example out of this PR and it could be the start of the cookbook or tutorial

sjrl · 2026-01-16T10:30:27Z

haystack_experimental/components/agents/agent.py

+from __future__ import annotations
+


Is this needed?

sjrl · 2026-01-16T10:30:52Z

haystack_experimental/components/agents/agent.py

+if TYPE_CHECKING:
+    from haystack_experimental.memory_stores.mem0.memory_store import Mem0MemoryStore


This can be dropped once this comment is addressed

sjrl · 2026-01-16T10:33:44Z

haystack_experimental/components/agents/agent.py

        result = {**exe_context.state.data}
        if msgs := result.get("messages"):
-            result["last_message"] = msgs[-1]
+            result["messages"] = msgs


This should be removed, it looks redundant. msgs is already equal to result["messages"]

sjrl · 2026-01-16T10:36:05Z

haystack_experimental/components/agents/agent.py

+            new_memories = [
+                message for message in msgs if message.role.value == "user" or message.role.value == "assistant"
+            ]
+            if self._memory_store:
+                self._memory_store.add_memories(messages=new_memories, **memory_store_kwargs)


Let's create the new list only if it will be used. Also I think it's simpler to just throw out the system message, since that's what contains the retrieved memories currently right?

Suggested change

new_memories = [

message for message in msgs if message.role.value == "user" or message.role.value == "assistant"

]

if self._memory_store:

self._memory_store.add_memories(messages=new_memories, **memory_store_kwargs)

if self._memory_store:

new_memories = [

message for message in msgs if message.role.value != "system"

]

self._memory_store.add_memories(messages=new_memories, **memory_store_kwargs)

sjrl · 2026-01-16T10:37:19Z

haystack_experimental/components/agents/agent.py

        if msgs := result.get("messages"):
-            result["last_message"] = msgs[-1]
+            result["messages"] = msgs
+            result["last_message"] = msgs[-1] if msgs else None


Revert back. The if msgs := result.get("messages"): already covers this. Using the walrus operator here means that if msgs is None then everything under the if statement is skipped

Suggested change

result["last_message"] = msgs[-1] if msgs else None

result["last_message"] = msgs[-1]

sjrl · 2026-01-16T10:44:25Z

@Amnah199 some remaining tasks:

update the readme with the new experiment
add the pydocs file for the memory store under pydoc
also apologies, but you'll need to do resolve some merge conflicts. Mostly to do with dropping python 3.9 support and my PR fix: Update experimental Agent to be compatible with Haystack 2.22.0 #423 which fixed some issues between haystack-experimental and haystack 2.22

sjrl · 2026-01-16T10:50:13Z

haystack_experimental/memory_stores/types/protocol.py

+        Search for memories in the store.
+
+        :param query: Text query to search for. If not provided, all memories may be returned.
+        :param filters: Backend-specific filter structure.


I think ideally we don't directly use the backend-specific filter structure but instead use the Haystack filter syntax. I think we should reuse the filter structure we already have here https://docs.haystack.deepset.ai/docs/metadata-filtering

So action items would-be:

update this docstring to indicate that Haystack filters are used. Look at the protocol for DocumentStore for inspiration

update the Mem0MemoryStore to convert Haystack filters into ones used by Mem0. You can find an example of this how we do this for our OpenSearch integration here. Basically the normalize_filters function converts Haystack filters to OpenSearch filters.

Thanks for the clear instructions!

sjrl · 2026-01-16T10:51:18Z

haystack_experimental/memory_stores/types/protocol.py

+        self,
+        *,
+        query: Optional[str] = None,


I get the impression that query should always be required right? Or is there a way to call Mem0 without a query to search for memories?

sjrl · 2026-01-16T10:53:42Z

haystack_experimental/memory_stores/types/protocol.py

+        user_id: Optional[str] = None,
+        run_id: Optional[str] = None,
+        agent_id: Optional[str] = None,
+        include_memory_metadata: bool = False,


Let's replace this with just kwargs

Suggested change

include_memory_metadata: bool = False,

**kwargs: Any,

this way any specific implementation of MemoryStore can add additional keyword arguments and still satisfy the protocol.

And the include_memory_metadata doesn't feel appropriate to enforce for a generic protocol.

sjrl · 2026-01-16T11:00:01Z

haystack_experimental/memory_stores/types/protocol.py

+    def add_memories(
+        self,
+        *,
+        messages: list[ChatMessage],
+        user_id: Optional[str] = None,
+        run_id: Optional[str] = None,
+        agent_id: Optional[str] = None,
+    ) -> list[str]:


I think a good middle of the road approach here would be to change this to

Suggested change

def add_memories(

self,

*,

messages: list[ChatMessage],

user_id: Optional[str] = None,

run_id: Optional[str] = None,

agent_id: Optional[str] = None,

) -> list[str]:

def add_memories(

self,

*,

messages: list[ChatMessage],

user_id: Optional[str] = None,

**kwargs: Any,

) -> list[str]:

so we always require an implementation of MemoryStore at least has a user_id but then all other ids that may be specific to the specific backend aren't required for all implementations.

If you agree let's update all functions in this protocol to work this way

sjrl · 2026-01-16T12:06:07Z

haystack_experimental/components/agents/agent.py

+
+        # Retrieve memories from the memory store
+        if self._memory_store:
+            retrieved_memory = self._memory_store.search_memories_as_single_message(


We should use the search_memories method here otherwise if we implement another backend MemoryStore it won't work with Agent since Agent relies on a method not defined in the protocol

Right, but I think one of the concerns was that we wanted to pass memories as a single message instead of a bunch of system messages for the model to process memories more efficiently.
I am thinking we can add this method to the protocol perhaps.

Right okay. Perhaps we can go back on what I said earlier and do the formatting of the single ChatMessage inside of agent here and drop the search_memories_as_single_message from the memory store

sjrl · 2026-01-16T12:13:01Z

haystack_experimental/memory_stores/mem0/memory_store.py

+            if include_memory_metadata:
+                # we also include the mem0 related metadata i.e. memory_id, score, etc.
+                # metadata
+                for memory in memories["results"]:
+                    meta = memory["metadata"].copy() if memory["metadata"] else {}
+                    meta["retrieved_memory_metadata"] = memory.copy()
+                    meta["retrieved_memory_metadata"].pop("memory")
+                    messages = [
+                        ChatMessage.from_system(text=memory["memory"], meta=meta) for memory in memories["results"]
+                    ]
+            else:
+                # we only include the metadata stored in the memory in ChatMessage
+                messages = [
+                    ChatMessage.from_system(text=memory["memory"], meta=memory["metadata"])
+                    for memory in memories["results"]
+                ]
+            return messages


We can simplify this to

Suggested change

if include_memory_metadata:

# we also include the mem0 related metadata i.e. memory_id, score, etc.

# metadata

for memory in memories["results"]:

meta = memory["metadata"].copy() if memory["metadata"] else {}

meta["retrieved_memory_metadata"] = memory.copy()

meta["retrieved_memory_metadata"].pop("memory")

messages = [

ChatMessage.from_system(text=memory["memory"], meta=meta) for memory in memories["results"]

]

else:

# we only include the metadata stored in the memory in ChatMessage

messages = [

ChatMessage.from_system(text=memory["memory"], meta=memory["metadata"])

for memory in memories["results"]

]

return messages

messages = []

for memory in memories["results"]:

meta = memory["metadata"].copy() if memory["metadata"] else {}

# we also include the mem0 related metadata i.e. memory_id, score, etc.

if include_memory_metadata:

meta["retrieved_memory_metadata"] = memory.copy()

meta["retrieved_memory_metadata"].pop("memory")

messages.append(

ChatMessage.from_system(text=memory["memory"], meta=meta)

)

return messages

sjrl · 2026-01-16T12:14:45Z

haystack_experimental/memory_stores/mem0/memory_store.py

+
+        try:
+            self.client.delete_all(**ids, **kwargs)
+            logger.info(f"All memories deleted successfully for scope {ids}")


Let's make sure to not use f-strings in log statements and use the expected Haystack format

Suggested change

logger.info(f"All memories deleted successfully for scope {ids}")

logger.info("All memories deleted successfully for scope {ids}", ids=ids)

sjrl · 2026-01-16T12:15:16Z

haystack_experimental/memory_stores/mem0/memory_store.py

+        """
+        try:
+            self.client.delete(memory_id=memory_id, **kwargs)
+            logger.info(f"Memory {memory_id} deleted successfully")


Suggested change

logger.info(f"Memory {memory_id} deleted successfully")

logger.info("Memory {memory_id} deleted successfully", memory_id=memory_id)

sjrl · 2026-01-16T12:17:52Z

haystack_experimental/memory_stores/mem0/memory_store.py

+    # mem0 doesn't allow passing filter to delete endpoint,
+    # we can delete all memories for a user by passing the user_id


It makes sense it doesn't all passing the filters otherwise it wouldn't be all right?

I think it's fine to remove this comment.

sjrl · 2026-01-16T12:20:10Z

haystack_experimental/memory_stores/types/protocol.py

+    def delete_all_memories(
+        self,
+        *,
+        user_id: Optional[str] = None,
+        run_id: Optional[str] = None,
+        agent_id: Optional[str] = None,
+    ) -> None:


I noticed that in your Mem0 implementation that you are exposing additional kwargs in this function so we should also add it to the protocol. Dropping run_id and agent_id based on this comment

Suggested change

def delete_all_memories(

self,

*,

user_id: Optional[str] = None,

run_id: Optional[str] = None,

agent_id: Optional[str] = None,

) -> None:

def delete_all_memories(

self,

*,

user_id: Optional[str] = None,

**kwargs: Any

) -> None:

sjrl · 2026-01-16T12:20:57Z

haystack_experimental/memory_stores/types/protocol.py

+        """
+        ...
+
+    def delete_memory(self, memory_id: str) -> None:


Same here I noticed that you expose kwargs in the Mem0 implementation so let's add it to the protocol

Suggested change

def delete_memory(self, memory_id: str) -> None:

def delete_memory(self, memory_id: str, **kwargs: Any) -> None:

sjrl · 2026-01-16T12:30:22Z

haystack_experimental/components/agents/agent.py

+            memory_text = f"Here are the relevant memories for the user's query: {retrieved_memory.text}"
+            updated_memory = ChatMessage.from_system(text=memory_text, meta=retrieved_memory.meta)
+        else:
+            updated_memory = None
+
+        combined_messages = messages + [updated_memory] if updated_memory else messages
+        if updated_system_prompt is not None:
+            combined_messages = [ChatMessage.from_system(updated_system_prompt)] + combined_messages


I'm not sure this is a good idea. Typically only one system prompt is ever used in the Chat History.

Sorry if we discussed this before, but which model providers have you tested that this works with?

If we really want to put the memories in a system prompt (I'm still not 100% this is best idea) we should probably add them directly to the updated_system_prompt so then at least we have a single system message.

Add mem0 integration

097aaaf

Amnah199 added 17 commits November 24, 2025 20:09

Add custom prompt

df96c56

Merge branch 'main' of https://github.com/deepset-ai/haystack-experim…

fd6288e

…ental into mem0-integration

Udpates

114d784

Update tests

7af23ae

Fix errors and update

56fe622

Fix licenses

8024c2d

Updates

69d3fbf

Fixes

4ca6164

Fix linting

18ebd92

Fix linting

73426b5

Updates

b063430

Update agent logic

9582257

Fix linting

758067b

fix linting

d6d503a

Fix linting

9a7dbf7

Update agent

b6ca11e

Update tests

0790f5b

Amnah199 marked this pull request as ready for review December 3, 2025 11:54

Amnah199 requested a review from a team as a code owner December 3, 2025 11:54

Amnah199 requested review from mpangrazzi and sjrl and removed request for a team and mpangrazzi December 3, 2025 11:54

sjrl reviewed Dec 3, 2025

View reviewed changes

haystack_experimental/components/memory_agents/agent.py Outdated Show resolved Hide resolved

sjrl reviewed Dec 3, 2025

View reviewed changes

haystack_experimental/components/memory_agents/agent.py Outdated Show resolved Hide resolved

sjrl reviewed Dec 3, 2025

View reviewed changes

haystack_experimental/components/memory_agents/agent.py Outdated Show resolved Hide resolved

sjrl reviewed Dec 3, 2025

View reviewed changes

haystack_experimental/components/memory_agents/agent.py Outdated Show resolved Hide resolved

sjrl reviewed Dec 3, 2025

View reviewed changes

haystack_experimental/memory_stores/examples/memory_store_with_agent.py Outdated Show resolved Hide resolved

Amnah199 added 4 commits December 9, 2025 16:05

Add new search method

c2fb710

Remove print statements

30b1525

Fix linting error

00313dd

Update tests

0dae67c

Amnah199 commented Dec 9, 2025

View reviewed changes

pyproject.toml Outdated Show resolved Hide resolved

Amnah199 added 3 commits December 9, 2025 18:11

Update the dependency

5e6eb0f

Test update

687688b

Fix bug

f79d8f0

Amnah199 requested a review from sjrl December 9, 2025 14:08

Amnah199 changed the title ~~Add Mem0 integration / Support memory in Agent~~ Add Mem0 integration - support for Mem0 platform Jan 15, 2026

sjrl reviewed Jan 16, 2026

View reviewed changes

haystack_experimental/components/agents/agent.py

Comment on lines +1 to +2

from __future__ import annotations

Copy link

Contributor

sjrl Jan 16, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Is this needed?

sjrl reviewed Jan 16, 2026

View reviewed changes

		if TYPE_CHECKING:
		from haystack_experimental.memory_stores.mem0.memory_store import Mem0MemoryStore

	result["last_message"] = msgs[-1] if msgs else None
	result["last_message"] = msgs[-1]

	logger.info(f"All memories deleted successfully for scope {ids}")
	logger.info("All memories deleted successfully for scope {ids}", ids=ids)

	logger.info(f"Memory {memory_id} deleted successfully")
	logger.info("Memory {memory_id} deleted successfully", memory_id=memory_id)

		# mem0 doesn't allow passing filter to delete endpoint,
		# we can delete all memories for a user by passing the user_id

	def delete_memory(self, memory_id: str) -> None:
	def delete_memory(self, memory_id: str, **kwargs: Any) -> None:

Add Mem0 integration - support for Mem0 platform #391

Are you sure you want to change the base?

Add Mem0 integration - support for Mem0 platform #391

Uh oh!

Conversation

Amnah199 commented Oct 28, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Related Issues

Proposed Changes:

Key Changes

How did you test it?

Notes for the reviewer

Checklist

Uh oh!

coveralls commented Oct 28, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Pull Request Test Coverage Report for Build 20066169239

Details

💛 - Coveralls

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

sjrl commented Jan 16, 2026

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

sjrl Jan 16, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Amnah199 commented Oct 28, 2025 •

edited

Loading

coveralls commented Oct 28, 2025 •

edited

Loading

sjrl Jan 16, 2026 •

edited

Loading