add() stops reporting memories the vector store never saved

If you run mem0 open source, the first fix in 2.2.1 is in add. Say add extracts three memories, and the vector store rejects the one about a penicillin allergy. Version 2.2.0 still returned all three as added, and wrote all three to history. Now only records that were actually inserted get a history entry, entity links, and a place in the results.

Slide text

Core LLMs Vector stores Upgrade

add() stops reporting memories the vector store never saved

2.2.0 add() returned

User's name is Aryan ADD

User is allergic to penicillin ADD

User works as an engineer ADD

2.2.1 add() returns

User's name is Aryan ADD

User is allergic to penicillin not stored

User works as an engineer ADD

The second core fix brings back a pattern some of you were already using. You can open a Memory in a with block again, and an AsyncMemory in an async with block. When the block ends, the instance is closed for you. Errors raised inside the block still reach your code.

Slide text

Core LLMs Vector stores Upgrade

Memory and AsyncMemory work as context managers again

mem0/memory/main.py

Python Node JS

with Memory() as m:
    ...

async with AsyncMemory() as m:
    ...

On exit, the instance closes. __exit__ calls self.close() and returns False, so an exception inside the block still propagates.

On AWS Bedrock, a Claude reasoning model can answer with a reasoning block first and the text block second. The Anthropic path always read the first block, so these calls failed with a KeyError. It now returns the first block that carries text. An empty content list returns an empty string.

Slide text

Core LLMs Vector stores Upgrade

Claude reasoning models on AWS Bedrock stop failing with KeyError

Converse response

Python Node JS

"content": [
  {"reasoningContent": {
      "reasoningText": {"text": "step by step..."}}},
  {"text": "final answer"}
]

2.2.0

content[0]["text"]

KeyError

2.2.1

first block with "text"

"final answer"

Turbopuffer users get two fixes. The first is filters. The store only read greater than or equal, and less than or equal. Every other operator was dropped silently, so a not equal or an in filter came back unfiltered. Now all eight apply: equal, not equal, greater than, less than, the two inclusive ones, in, and not in.

Slide text

Core LLMs Vector stores Upgrade

Turbopuffer filters now apply all eight operators

2.2.0

eq ne gt gte lt lte in nin

Dropped without an error, so the query returned unfiltered results.

2.2.1

eq ne gt gte lt lte in nin

Mapped to Turbopuffer's Eq, NotEq, Gt, Gte, Lt, Lte, In and NotIn.

The second fix is search scores. Turbopuffer and S3 Vectors both turned a distance into a score as one minus the distance, which only works for cosine. With euclidean distances, Turbopuffer scores went negative past a distance of 1 and inverted ranking, and S3 Vectors collapsed most scores to zero. Both now use one over one plus the distance, which stays between zero and one.

Slide text

Core LLMs Vector stores Upgrade

Euclidean search scores on Turbopuffer and S3 Vectors stay between 0 and 1

Distance	Turbopuffer 2.2.01 − d	S3 Vectors 2.2.0max(0, 1 − d)	Both in 2.2.11 / (1 + d)
0.5	0.5	0.5	0.67
2	−1	0	0.33
4	−3	0	0.2

Worked example from the formulas. Applies to euclidean_squared on Turbopuffer and euclidean on S3 Vectors; cosine scores are unchanged.

Two of these fixes can raise an error where your code used to carry on. If the vector store rejects every extracted memory, add now raises a VectorStoreError instead of returning memories that were never stored. Your messages are still saved for a retry. And a Turbopuffer filter with an unsupported operator raises a value error.

Slide text

Core LLMs Vector stores Upgrade

Two calls now raise errors instead of carrying on

VectorStoreError

Memory.add() and AsyncMemory.add()

When none of the extracted memories could be inserted. The raw messages are still saved, so a later retry can re-extract them.

ValueError

Turbopuffer filters

When a filter uses an operator outside the eight. 2.2.0 dropped it and returned unfiltered results.

Upgrading is one command. Install mem0ai 2.2.1 from pip. Before you upgrade, check any code that assumes add always returns, and wrap it to catch the new error. After that, add stops reporting memories the vector store never saved.

Slide text

Core LLMs Vector stores Upgrade

Upgrade to mem0ai 2.2.1, and catch VectorStoreError around add()

$pip install --upgrade mem0ai==2.2.1

Hold off only if your code needs to handle a failed write in add() first, since it now raises instead of returning.

Imported from mem0.exceptions, the same module as LLMError.