Mem0 v2.2.1 stops reporting memories it never stored

In this Mem0 release, the add call changes what it reports. Until now, a record the vector store rejected still came back as a successful add event, although it was never stored.

Slide text

Mem0 v2.2.1 stops reporting memories it never stored

add() now returns only what the vector store accepted.

docs/changelog/sdk.mdx

What the add call returns.

Now only the records the vector store actually inserted count. Those are the ones written to history, linked to entities, and returned to your code, in both Memory and AsyncMemory.

If none of the extracted memories could be inserted, the add call now raises a VectorStoreError. Before, it returned those memories anyway, so your code carried on as if they were saved.

The context manager protocol is back on Memory and AsyncMemory. Open one in a with block, or an async with block, and the instance closes itself when the block exits.

with Memory() as m:
    ...

async with AsyncMemory() as m:
    ...

Vector store search.

Turbopuffer filters now apply every operator, in both the Python and TypeScript SDKs. Before, only the range operators were read. Equals, not equals, in and not in were dropped silently, and the query came back unfiltered.

Turbopuffer search scores now respect the distance metric. With squared euclidean distance, the old score went negative for any distance above one, which turned the ranking upside down. Now it stays between zero and one.

S3 Vectors search scores are now metric-aware. With euclidean distance, the old formula collapsed most scores to zero. Now every distance maps to its own score, and a closer match always scores higher.

Models and providers.

On AWS Bedrock, Claude reasoning models can send a reasoning block before the text. Mem0 used to take only the first block and fail with a key error. Now it returns the first block that carries text.

In the TypeScript SDK, a DeepSeek or xAI config without its own base URL or model used to point at OpenAI, with an OpenAI model name. Now it falls back to that provider's own defaults.

Installing from npm.

Installing mem0ai from npm no longer pulls in the Postgres driver. It's an optional peer dependency now, with a caret range in place of the exact pin, so it stops clashing with your app's own version.

If you use the PGVector store, the SDK now loads the Postgres driver only when that store is used. So install it next to mem0ai with one npm command.

npm install pg

Mem0 for Hermes Agent.

Mem0's memory for Hermes Agent, from Nous Research, now has its own plugin. Install it from the Mem0 repository, enable it, and run the memory setup wizard. It recalls memories before each reply.

hermes plugins install \
  mem0ai/mem0/integrations/hermes-plugin-mem0
hermes plugins enable mem0
hermes memory setup mem0

The setup wizard offers three backends. Mem0 Cloud with an API key, your own self-hosted Mem0 server, or the open source SDK running inside Hermes with your own model and vector store.

Both SDKs shipped on September twenty-fifth. Upgrade mem0ai from PyPI for Python, or from npm for TypeScript.

Slide text

Mem0 v2.2.1

Python SDK 2.2.1 and TypeScript SDK 3.3.1, 25 September 2026

pip install -U mem0ai
npm install mem0ai@latest

pyproject.toml · mem0-ts/package.json