feat: Implement memory optimization #5229

uzaxirr · 2025-10-29T06:13:59Z

Summary

Describe key changes, mention related issues or motivation for the changes.

3 main classes:

MemoryOptimizationStrategyType (Enum)

catalog of available strategies
Prevents typos and misunderstanding giving a better DX
from_string() method for backward compatibility allows to use stratergies as simple string Eg: stratergy="merge" as well. This gives good DX

MemoryOptimizationStrategyType.SUMMARIZE 
MemoryOptimizationStrategyType.MERGE

MemoryOptimizationStrategy (Abstract Base Class)

Defines contract for all strategies
Provides shared utilities (count_tokens())

MemoryOptimizationStrategyFactory

Centralized strategy instantiation
similar to ChunkingStrategyFactory

Users can also define their custom strategy.

Type of change

Checklist

Code complies with style guidelines
Ran format/validation scripts (./scripts/format.sh and ./scripts/validate.sh)
Self-review completed
Documentation updated (comments, docstrings)
Examples and guides: Relevant cookbook examples have been included or updated (if applicable)
Tested in clean environment
Tests added/updated (if applicable)

Additional Notes

Add any important context (deployment instructions, screenshots, security considerations, etc.)

dirkbrnd · 2025-11-04T04:10:08Z

libs/agno/agno/memory/strategy.py

+        raise ValueError(f"Unsupported optimization strategy: {strategy_name}. Valid options: {[e.value for e in cls]}")
+
+
+class MemoryOptimizationStrategy(ABC):


Base class should be in strategies.

dirkbrnd · 2025-11-04T04:10:36Z

libs/agno/agno/memory/strategy.py

+class MemoryOptimizationStrategy(ABC):
+    """Base class for memory optimization strategies."""
+
+    def count_tokens(self, memories: List["UserMemory"]) -> int:


This should be a general utility function. We can use it in many places.

dirkbrnd · 2025-11-04T04:11:14Z

libs/agno/agno/memory/strategies/merge.py

+from agno.utils.log import log_debug
+
+
+class MergeStrategy(MemoryOptimizationStrategy):


This is probably the only strategy that makes sense for now. Replace all memories with a single memory.

dirkbrnd · 2025-11-04T04:12:19Z

libs/agno/agno/memory/strategies/merge.py

+
+        # Concatenate input and feedback fields
+        input_parts = [mem.input for mem in memories if mem.input]
+        merged_input = "\n---\n".join(input_parts) if input_parts else None


I think that doesn't make sense to keep. The input kinda falls away if you optimize

dirkbrnd · 2025-11-04T04:12:38Z

libs/agno/agno/memory/strategies/merge.py

+        merged_input = "\n---\n".join(input_parts) if input_parts else None
+
+        feedback_parts = [mem.feedback for mem in memories if mem.feedback]
+        merged_feedback = "\n---\n".join(feedback_parts) if feedback_parts else None


Same with feedback (although we don't use it)

dirkbrnd · 2025-11-04T04:12:57Z

libs/agno/agno/memory/strategies/merge.py

+        feedback_parts = [mem.feedback for mem in memories if mem.feedback]
+        merged_feedback = "\n---\n".join(feedback_parts) if feedback_parts else None
+
+        # Check if agent_id and team_id are consistent


These as well. Since a whole new memory is created the original memory creator is lost

dirkbrnd · 2025-11-04T04:14:33Z

libs/agno/agno/memory/manager.py

+        self,
+        user_id: Optional[str] = None,
+        token_limit: int = 2000,
+        strategy: Union[str, MemoryOptimizationStrategyType, MemoryOptimizationStrategy] = "summarize",


We should only support MemoryOptimizationStrategyType and MemoryOptimizationStrategy

dirkbrnd · 2025-11-04T04:15:29Z

libs/agno/agno/memory/manager.py

+
+        Args:
+            user_id: User ID to optimize memories for. Defaults to "default".
+            token_limit: Maximum tokens for all memories combined.


This doesn't work for summarize, because there the token limit is per memory?

In general I don't think we should have token limit, it doesn't really make sense. The Model should just summarize.

dirkbrnd · 2025-11-04T04:15:58Z

libs/agno/agno/memory/manager.py

+                - String: "summarize", "merge"
+                - Enum: MemoryOptimizationStrategyType.SUMMARIZE, MemoryOptimizationStrategyType.MERGE
+                - Instance: Custom MemoryOptimizationStrategy instance
+            model: Model to use for optimization. Defaults to manager's model.


I don't think this should be a parameter. You should just create the memory manager with a model and call optimize.

dirkbrnd · 2025-11-04T04:16:38Z

libs/agno/agno/memory/manager.py

+
+        # Get strategy instance
+        if isinstance(strategy, str):
+            strategy_type = MemoryOptimizationStrategyType.from_string(strategy)


Don't need a from_string, it is an enum so MemoryOptimizationStrategyType[strategy] should work

dirkbrnd · 2025-11-04T04:17:50Z

libs/agno/agno/memory/manager.py

+        current_tokens = strategy_instance.count_tokens(memories)
+        log_debug(f"Current token count: {current_tokens}, Target: {token_limit}")
+
+        if current_tokens <= token_limit:


I think this is a bit overengineerd. Let the user decide themselves when they want to optimize. We can give them the util for counting the tokens if needed.

dirkbrnd · 2025-11-04T04:18:46Z

libs/agno/agno/memory/manager.py

+            for opt_mem in optimized_memories:
+                if opt_mem.memory_id:
+                    # Update existing memory
+                    self.replace_user_memory(


I believe replace is upsert, so we should be able to just call that for all cases.

dirkbrnd · 2025-11-04T04:19:13Z

libs/agno/agno/memory/strategy.py

+        raise NotImplementedError
+
+    @abstractmethod
+    def optimize(


And aoptimize

libs/agno/agno/memory/strategies/summarize.py

dirkbrnd · 2025-11-05T19:35:47Z

libs/agno/agno/memory/strategy.py

I think this should move to the init of the strategies directory

dirkbrnd · 2025-11-05T19:36:06Z

libs/agno/agno/memory/strategies/summarize.py

+from uuid import uuid4
+
+from agno.db.schemas import UserMemory
+from agno.memory.strategies.base import MemoryOptimizationStrategy


There is no base?

dirkbrnd · 2025-11-05T19:36:40Z

libs/agno/agno/memory/strategies/summarize.py

+        self,
+        memories: List[UserMemory],
+        model: Model,
+        user_id: Optional[str] = None,


I don't think this needs to be a parameter. You can just grab it from the first memory?

dirkbrnd · 2025-11-05T19:37:41Z

libs/agno/agno/memory/manager.py

+        strategy: Union[
+            MemoryOptimizationStrategyType, MemoryOptimizationStrategy
+        ] = MemoryOptimizationStrategyType.SUMMARIZE,
+        apply: bool = False,


I think this should be true by default. That seems like the default behaviour and that a "dry run" should be opt-in?

dirkbrnd · 2025-11-05T19:41:36Z

libs/agno/agno/memory/manager.py

+                mem for mem in memories if mem.memory_id and mem.memory_id not in optimized_memory_ids
+            ]
+
+            # Delete removed memories


Shouldn't we rather just clear all the user's memories and add the new ones? So it is a complete replace.

This is doing the same thing. It's looping through all the user's memory and then deleting them. Do you want me to call any specific function which already pre-exists?

uzaxirr added 2 commits October 29, 2025 11:25

add optimise_memories

7e8246c

rm emj

855894a

uzaxirr requested a review from a team as a code owner October 29, 2025 06:14

Merge branch 'main' into mem-opt

42af1c0

uzaxirr self-assigned this Oct 29, 2025

uzaxirr and others added 3 commits October 29, 2025 14:37

Merge branch 'main' into mem-opt

06dbe10

Merge branch 'main' into mem-opt

f814dd7

Merge branch 'main' into mem-opt

64b87ad

dirkbrnd reviewed Nov 4, 2025

View reviewed changes

libs/agno/agno/memory/strategy.py Outdated

raise NotImplementedError

@abstractmethod

def optimize(

Copy link

Contributor

dirkbrnd Nov 4, 2025

Choose a reason for hiding this comment

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

And aoptimize

dirkbrnd reviewed Nov 4, 2025

View reviewed changes

libs/agno/agno/memory/strategies/summarize.py Show resolved Hide resolved

uzaxirr and others added 7 commits November 4, 2025 18:05

Merge branch 'main' into mem-opt

03cf572

Merge branch 'main' into mem-opt

b10e78f

add comments

32558fd

rename merge to summarize

6c293ab

update cookbook

5d2ebb7

Merge branch 'main' into mem-opt

8bf1546

Merge branch 'main' into mem-opt

fd3f3d0

dirkbrnd reviewed Nov 5, 2025

View reviewed changes

libs/agno/agno/memory/strategy.py

Copy link

Contributor

dirkbrnd Nov 5, 2025

Choose a reason for hiding this comment

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

I think this should move to the init of the strategies directory

dirkbrnd reviewed Nov 5, 2025

View reviewed changes

uzaxirr and others added 4 commits November 6, 2025 01:14

add base class

689481d

cmnts

cf2ee84

Merge branch 'main' into mem-opt

01ed8ae

format

92d45a6

		raise ValueError(f"Unsupported optimization strategy: {strategy_name}. Valid options: {[e.value for e in cls]}")


		class MemoryOptimizationStrategy(ABC):

		from agno.utils.log import log_debug


		class MergeStrategy(MemoryOptimizationStrategy):

feat: Implement memory optimization #5229

Are you sure you want to change the base?

feat: Implement memory optimization #5229

Uh oh!

Conversation

uzaxirr commented Oct 29, 2025

Summary

Type of change

Checklist

Additional Notes

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!

Choose a reason for hiding this comment

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!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants