LoginSign Up
← Back to all sparks
M

Mem0

AI-ASSISTANTS
Velocity5.0

Memory layer for AI agents and personalized LLM applications

mem0.ai ↗

Mem0's release stream is provider breadth on one side and filter correctness on the other

agent-memoryvector-storessdkmulti-tenancyintegrationsretrieval
◆Current state
Mem0 is a memory layer for AI agents, released as separate artifacts per surface — Python SDK, Node SDK, an n8n integration node, an OpenCode plugin — each tagged independently. The recent window splits cleanly in two: new storage and embedding providers arriving (Oracle AI Vector Search with HNSW/IVF indexes and six distance metrics, AWS Bedrock embeddings), and a run of fixes to filter translation across Chroma, Cassandra, OpenSearch and Qdrant. Packaging work also completed a lazy-loading migration so importing the OSS package no longer pulls in every provider SDK.
◆Where it's heading
The filter bugs are the more revealing half. Chroma where-clauses were dropping conditions, Cassandra compound filters stopped after the first operator, an OpenSearch wildcard matched literally — each one silently widened or emptied a result set rather than failing loudly. Fixing that cluster, alongside making user_id, agent_id and run_id immutable after creation, is a memory layer hardening its retrieval and tenancy guarantees at the point where a wrong answer is invisible. Provider breadth continues in parallel, but correctness is where the recent effort concentrates.
◆Prediction
Expect the provider matrix to keep widening while filter-translation parity across stores continues to be squared off; the per-surface release split suggests the n8n and editor-plugin integrations will keep versioning on their own cadence.

◆Recent moves

  1. 1d ago

    n8n node republished with signed npm provenance

    A republish through GitHub Actions purely to attach a signed provenance attestation, which the manual 0.1.0 upload lacked and which n8n Creator Portal verification requires. No functional change.

    View source ↗
  2. 6d ago

    Node SDK fixes dropped filter conditions in Chroma and Cassandra

    Cassandra compound filters stopped applying operators after the first, and the Chroma where-clause translator collapsed same-field ranges and negated conditions. Both silently widened searches rather than erroring, which is the worst failure mode for a memory layer.

    View source ↗
  3. 6d ago

    Python SDK adds an Oracle AI Vector Search provider

    Oracle AI Vector Search joins the store matrix with connection pooling, HNSW and IVF indexes, JSON metadata filtering and six distance metrics. Shipped alongside a fix for OpenSearch wildcard filters that previously matched the literal string.

    View source ↗
  4. 8d ago

    Python SDK makes memory identity fields immutable after creation

    update() metadata can no longer overwrite or inject user_id, agent_id, run_id or actor_id, closing a path that could move a memory into another tenant's scope. Also fixes reset() leaving stale vectors on on-disk Qdrant when the collection directory can't be removed.

    View source ↗
  5. 8d ago

    OpenCode plugin reads its API key from shell profiles

    Falls back to reading MEM0_API_KEY from .zshrc, .bashrc and similar when it isn't exported into the process environment, fixing startup on desktop clients launched outside a shell.

    View source ↗
  6. 8d ago

    Node SDK adds Bedrock embeddings and finishes lazy provider loading

    An AWS Bedrock embedding provider lands, and the lazy-loading work completes so importing the OSS package no longer requires every LLM, embedder and vector store SDK to be installed. A real reduction in install weight for anyone using one provider.

    View source ↗