Mem0 changelog: what's new each month
Every stable Mem0 release summarised by month: the highlights, new features, improvements, fixes and anything you need to act on. 4 months covered; the current month updates daily.
September 2026
23 releases: 2.0.20 → 3.3.1
September brought major updates across Mem0's plugin ecosystem, with streamlined memory search behavior, new user profile capabilities, and improved platform attribution and telemetry. Multiple integration plugins were refactored to share core utilities while removing deprecated Dream consolidation features.
- Removed Dream consolidation, automatic scheduling, and Dream configuration from OpenClaw, DeepSeek, OpenCode, and Pi integrations; existing configurations using these features must be updated.
- Removed OpenClaw's separate 2,000-character extraction cutoff; full redacted text now retained for selected user and assistant messages.
- Global memory tool scope now requires explicit user enablement; empty and wildcard identities are rejected.
- PostHog telemetry source tag changed from literal "plugin" to OPENCODEPLUGIN; saved insights filtering on source="plugin" will stop matching new events.
Highlights
- Added User Profiles API with methods to get, generate, and manage structured user profile data across MemoryClient and AsyncMemoryClient.
- Simplified memory search behavior across all plugins to use single searches instead of parallel searches, reducing unnecessary calls.
- Implemented surface-identity headers (X-Mem0-Source, X-Application, X-Mem0-Client) across Python, TypeScript, and Go SDKs for better platform call attribution.
- Fixed telemetry event duplication and loss issues across plugin ecosystem with improved event attribution and delivery.
- Refactored OpenClaw, DeepSeek, OpenCode, and Pi integrations to share core conversation preparation, redaction, and telemetry utilities.
New
- User Profiles API with getProfile(), generateProfile(), getProfileSettings(), updateProfileSettings(), sampleProfiles(), and getProfileJob() methods.
- Automatic memory recall during system-prompt assembly using the latest human prompt without repeated context injection.
- Automatic memory capture from durable session/event stream after completed turns.
- Global memory tool scope requiring explicit user enablement in plugin settings.
- Memory update and delete now accept mem0:<uuid> and [mem0:<uuid>] citations from tool results in addition to raw IDs.
- Install-count deduplication using keyFingerprint in config schema to count installs once per key.
Improved
- Vector store timestamp handling now guards against None values in Valkey insert() and update() paths.
- AWS Bedrock Anthropic integration now returns the first Converse content block carrying text instead of always reading content[0], supporting Claude reasoning models.
- Turbopuffer vector store now applies all filter operators (eq, ne, gt, gte, lt, lte, in, nin) instead of silently dropping non-range operators.
- Turbopuffer and S3 Vectors distance-to-score mapping now respects the configured distance metric for accurate scoring.
- OSS configuration no longer injects OpenAI defaults into non-OpenAI LLM providers like DeepSeek and xAI.
- Plugin telemetry now attributes each event to the plugin that produced it with updated pluginversion field.
- Memory capture preparation now filters conversation roles and redacts sensitive content.
- SDK client stack now bounded by dropping whole entries rather than slicing characters to preserve identifiers.
Fixed
- Memory add() no longer reports vector-store-rejected records as successful, only actually inserted records are written to history.
- Context-manager protocol restored on Memory and AsyncMemory classes to properly close instances on exit.
- Telemetry events no longer delivered twice and no longer lose parked events on flush.
- Turbopuffer search scores with euclideansquared metric no longer go negative and invert ranking.
- S3 Vectors search scores with euclidean metric no longer collapse most scores to zero.
- OpenClaw status command no longer crashes on unconfigured installations.
August 2026
27 releases: 2.0.15 → 0.1.1
August brought stability improvements, security patches, and expanded platform capabilities across Mem0's SDKs and integrations. The month focused on fixing data consistency issues, adding new memory filtering options, and introducing support for Oracle as a vector store.
- add() metadata no longer sets or overwrites memory identity scope; userid, agentid, runid, and actorid are stripped before processing.
- CLI: --categories flag removed from add command; use --custom-categories instead as the endpoint does not accept categories field.
- LLMReranker and RerankerFactory default model changed from gpt-4o-mini to gpt-5-mini.
- Ollama embedder import now raises ImportError instead of blocking on pip install; manually install ollama package.
Highlights
- Added Oracle AI Vector Search as a new vector store option with pooled connections and JSON filtering.
- New agent-scoped custom instructions allow separate extraction rules for agent-specific memories.
- Expanded search and memory options with reference dates, latest-only filtering, and keyword search across multiple SDK clients.
- Fixed critical data loss bug where deleteAll() could silently leave memories behind in accounts exceeding vector store page limits.
- Patched 32+ high and 57+ medium severity dependency vulnerabilities across the entire workspace.
New
- Oracle AI Vector Search vector store with HNSW/IVF indexes, optional indexAccuracy targeting, and JSON payload filtering.
- Agent custom instructions field (agentCustomInstructions / agentcustominstructions) for extraction rules applying only to agent-scoped memories.
- Qdrant server-side BM25 keyword search without client-side dependency (requires Qdrant >= 1.15.2).
- New CLI flags: --custom-instructions, --custom-categories, --structured-data-schema, --timestamp, --show-expired, --reference-date, --latest-only, --delete-linked.
- Deepseek plugin for automatic memory recall and server-side extraction in Deepseek frameworks.
- Strands integration with automatic recall injection and non-blocking client construction.
- Reference date, latest-only, and keyword search options in SearchMemoryOptions across Python and TypeScript clients.
- N8n community node with themed icons and MIT license for Creator Portal compatibility.
Improved
- deleteAll() now paginates through vector stores in batches of 1000 to handle accounts with large memory counts.
- Supabase list() pagination now works past PostgREST's 1000-row cap; search() warns when results may be truncated.
- Elasticsearch KNN search now respects topk parameter instead of capping at default 10 results.
- HuggingFace embedder now honors configured API keys and falls back to env vars when custom base URL is set.
- Redis and Valkey payloads now preserve identity fields (agentid, runid, userid) in snakecase format.
- Client identity calls removed from request critical path, speeding up first method calls on each client instance.
- Ollama embedder now raises clear ImportError instead of blocking on interactive pip install prompts.
- Multiple LLM and embedding providers now properly honor environment variable fallbacks for API keys and base URLs.
Fixed
- Fixed PGVector filter validation that allowed string values to be iterated character-by-character instead of treated as arrays.
- Fixed deleteUsers() performance by using shared fetch-based helper instead of axios with keepAlive: false.
- Fixed Redis search/list building invalid empty RediSearch filter expressions when filters contained only null/undefined values.
- Fixed Oracle connection/pool not being closed when initialization fails, preventing resource leaks.
- Fixed percent-escaping in userid/agentid/runid for session scope keys to prevent ambiguous key generation.
- Fixed createProceduralMemory() raising cryptic errors when LLM returns no content for summary.
July 2026
17 releases: 2.0.11 → 0.1.1
July brought major expansions to Mem0's vector store and LLM support, alongside critical security fixes and improved memory retrieval. The month also saw the completion of lazy-loading for optional dependencies and numerous fixes for edge cases across filters, entity handling, and identity field management.
- Memory.update() parameter data is deprecated in favor of text
- Client retrievalCriteria field removed from PromptUpdatePayload
- Client retrievalcriteria parameter removed from MemoryClient.updateproject() and Project.update()
- Update metadata can no longer set userid, agentid, runid, or actorid fields; these are now immutable after creation
Highlights
- Added 15+ new vector store providers including Oracle AI Vector Search, Pinecone, Weaviate, Milvus, MongoDB, and Elasticsearch.
- Added reranking to the OSS SDK with four provider options and per-search rerank control.
- Secured multiple vector stores and graph databases against injection attacks via filter validation and escaping.
- Improved memory retrieval with guaranteed prefetching and multi-hop search support.
- Completed lazy-loading of optional provider SDKs to reduce import overhead.
New
- Reranking for OSS with Cohere, ZeroEntropy, cross-encoder, and LLM-based providers
- Oracle AI Vector Search vector store with connection pooling and six distance metrics
- Fifteen new vector store providers: Pinecone, Weaviate, Milvus, Chroma, MongoDB, Elasticsearch, OpenSearch, Databricks, AWS Neptune Analytics, S3 Vectors, Azure MySQL, Google Vertex AI Vector Search, Turbopuffer, Upstash Vector, Valkey, Cassandra, Baidu Mochow
- Five new LLM providers: AWS Bedrock, xAI Grok, Together, vLLM, and Sarvam
- Four new embedding providers: Vertex AI, HuggingFace, FastEmbed, and Together
- Guaranteed retrieval with memory prefetching injected before agent execution
- Multi-hop retrieval with proactive multi-query search in memory policy
- Shell profile API key recovery from .zshrc, .bashrc, .zprofile, .bashprofile, or .profile
Improved
- Sharper tool descriptions for all memory tools explaining when to use each one
- Memory.update() now accepts text parameter; data parameter deprecated but still works
- Non-string entity IDs coerced to strings instead of crashing
- Lazy-loading of all optional provider SDKs to reduce import dependencies
- Dynamic URL path segments properly encoded for IDs with special characters
- OpenSearch filter handling improved to translate empty strings to exists queries
- Milvus wildcard search and BM25 text field handling improved
- Anthropic model sampling parameter detection automatic per model family and version
Fixed
- Embeddings no longer crash on OpenAI and Azure OpenAI batch count mismatch
- LLM extraction failures now re-raised instead of silently returning empty results
- Entity delete operations now return all deleted entities instead of only the last one
- Output formatters handle null memory fields gracefully
- Pinecone namespace-scoped delete and reset no longer wipe other namespaces
- Chroma where-clause filters no longer drop conditions on same-field ranges and multi-field $or operations
June 2026
13 releases: 1.0.13 → 3.0.12
June brought expanded LLM provider support, critical security updates, and major architectural improvements to the OpenCode plugin including native tool registration and automatic memory consolidation. Vector store reliability and memory handling were significantly enhanced across the platform.
- OpenCode plugin: Skills no longer copied into .opencode/ directory; installSkills() and cli.ts installer removed. Skills load via skills.paths configuration.
- OpenCode plugin: Ten skills removed (import, export, memory-reviewer, mem0, list-projects, stats, and five others). Nine focused skills remain.
- OpenCode plugin: session-metadata registration mode now returns early and registers only CLI commands, avoiding backend initialization.
Highlights
- Added Gemini via Vertex AI and MiniMax as new LLM providers, plus LiteLLM for proxy routing.
- OpenCode plugin now registers memory tools natively and loads skills in place via config instead of copying files.
- Introduced auto-dream for automatic memory consolidation and memory scope control (project, session, or global).
- Fixed critical crashes in vision message parsing, vector store operations, and async memory deletion.
- Remediated multiple CVEs including axios, undici, and LangChain community vulnerabilities.
New
- Gemini via Vertex AI as LLM provider option.
- MiniMax LLM provider backed by the OpenAI-compatible MiniMax API.
- LiteLLM provider for routing requests through local or hosted LiteLLM proxy.
- Memory scope parameter — search and manage memories at project, session, or global level.
- Auto-dream automatic memory consolidation gated by time, session count, and memory count thresholds.
- File-context injection in OpenCode plugin — injects prior work as system context before file reads.
- Session summary memory with structured type and source metadata for backend AI extraction.
- Contextual OSS-to-Platform notices surfacing adoption guidance and performance alerts.
Improved
- Native batched embedding support added to OllamaEmbedding, LM Studio, Together, HuggingFace, Vertex AI, and Google GenAI embedders.
- OpenCode plugin skills now load via config hook instead of filesystem copying.
- SessionStart activity timeline now shows type icons and relative age indicators for quick scanning.
- PGVector now supports connection via connectionString and ssl options as alternative to individual host/user/password/port.
- Anthropic tool choice now correctly formatted as object instead of bare string.
- Custom metadata fields now preserved during memory updates instead of being dropped.
- PostHog telemetry implemented as anonymous, fire-and-forget events; API key no longer visible in process list.
- Qdrant vector store now exposes https option for explicit TLS endpoint targeting.
Fixed
- Memory crash in parseVisionMessages when vision support is disabled or imageurl entries are malformed.
- Vector store search crashes in LangChain when result scores are None.
- AsyncMemory.reset() not resetting the entity store and leaving stale message history.
- Async deleteAll aborting on first error and leaving partial deletions; race condition corrupting linkedMemoryIds.
- PGVector and Redis insert/update crashes on entity payloads missing hash or timestamp fields.
- Valkey FT.SEARCH tag filter values allowing wildcard and operator injection through tenant-isolation filters.
Summaries are written automatically from the official release notes (full changelog ↗); check the original notes before relying on a detail. Mem0: pricing, features and alternatives · All changelogs