MemGPT
AI & LLMsLetta is the platform for building stateful agents: open AI with advanced memory that can learn and self-improve over time.
Release History
0.16.5This release bumps the version to 0.16.5. It appears to be a maintenance or chore release.
0.16.41 fixThis release primarily consists of maintenance updates, including updating GitHub templates and bumping the version from 0.16.2 to 0.16.4.
0.16.2This release primarily focuses on documentation updates, including corrections to the README and contributing guides, and bumps the version to 0.16.2.
0.16.11 fixThis patch release (v0.16.1) primarily fixes an issue with the provider name configuration for openai-proxy in LLMConfig.
0.16.01 fixThis release bumps the version to v0.16.0 and includes an update to the readme and architecture-specific OTEL installation logic.
0.15.1This release bumps the version to 0.15.1.
0.15.01 featureVersion 0.15.0 introduces context window support for grok-4 models. This release focuses on expanding model capabilities.
0.14.0This release bumps the version number to 0.14.0.
0.13.01 featureVersion 0.13.0 introduces Haiku 4.5 as a new reasoning model and includes documentation cleanup.
0.12.1Breaking2 fixes11 featuresThis release introduces the major `letta_v1_agent` architecture, offering broader provider compatibility and simpler control flow, alongside new features like Human-in-the-Loop and Parallel Tool Calling. It also deprecates and renames several older API endpoints.
0.12.0No release notes provided.
0.11.72 fixes17 featuresThis release introduces significant Human-in-the-Loop (HITL) support for tool execution and upgrades the Agent File schema to v2. It also brings substantial improvements to archival memory search capabilities and enhances model support for GPT-5, DeepSeek, and Anthropic.
0.11.6No release notes provided.
0.11.51 featureThis release introduces background mode support for message streaming, enhancing asynchronous operations.
0.11.42 featuresThis release deprecates legacy paths for azure and together, introduces an asyncio shield for stream timeouts, and adds step metrics recording.
0.11.31 featureVersion 0.11.3 primarily involves refactoring by moving dictconfig out of the getlogger function.
0.11.22 fixes1 featureThis release introduces a new max_steps parameter for agent export and fixes two bugs related to the Ollama provider, specifically concerning the embeddings endpoint URL and model type returns.
0.11.1Breaking1 fix2 featuresThis release introduces support for new LLM models like Claude Opus 4.1 and GPT-5, and enhances built-in tools by improving memory reliability and paginating the file grep tool.
0.11.0Breaking1 fix2 featuresThis release fully removes legacy clients, introduces Signoz tracing integration, optimizes Jinja templating performance, and raises the minimum required Python version to 3.11.
0.10.0Breaking5 fixes8 featuresThis release introduces the `LettaPing` message for stable long streaming connections and adds support for OAuth MCP providers. It also defaults agents to the new `memgpt_v2_agent` architecture and includes various performance improvements and bug fixes.
0.9.11 fix1 featureThis release focuses on bug fixes for the filesystem feature and introduces asynchronous rendering for jinja templates in core routes, alongside adding an agent tag reverse index.
0.9.03 featuresVersion 0.9.0 introduces the Letta Filesystem for enhanced document management and context control, along with new options for document parsing via Mistral OCR.
0.8.17This release bumps the version to v0.8.17, primarily consisting of maintenance chores.
0.8.16This release bumps the version to v0.8.16, primarily consisting of maintenance chores.
0.8.15This release bumps the version to v0.8.15, primarily consisting of maintenance chores.
0.8.14This release bumps the version number to 0.8.14.
0.8.13This release bumps the version number from 0.8.12 to 0.8.13.
0.8.12This release bumps the version to 0.8.12 and adds tracing capabilities to the polling mechanism.
0.8.11This release bumps the internal letta version to 0.8.11.
0.8.10This release (0.8.10) is a maintenance update, likely containing minor fixes or internal changes, as indicated by the 'chore' tag.
0.8.91 fix4 featuresThis release introduces support for different summarization providers, agent loop cancellation, and MCP custom headers, alongside various bug fixes.
0.8.83 fixes2 featuresThis release introduces new Feedback APIs for steps and includes several bug fixes, such as improvements to Anthropic streaming and agent timezone updates.
0.8.71 fix3 featuresThis release introduces new react and workflow agents, adds step feedback, and includes fixes for batch agents. Some existing tools have also been deprecated.
0.8.61 fix1 featureThis release introduces custom timezone support and fixes a bug in the bedrock integration.
0.8.5No release notes provided.
0.8.42 fixes2 featuresThis release introduces streamable HTTP support and authenticated MCP support, alongside fixes for OpenAI o-series models and Anthropic streaming.
0.8.32 fixesThis release includes fixes for issues found in the stdio and sse MCP connectors.
0.8.2No release notes provided.
0.8.1No release notes provided.
0.8.02 fixesThis release includes fixes for Ollama integration and token counting logic, alongside the official deprecation of legacy clients.
0.7.292 featuresVersion 0.7.29 introduces new configuration options for batch size and lookback settings.
0.7.281 featureVersion 0.7.28 introduces performance improvements. This release primarily focuses on internal optimizations.
0.7.271 fixThis release bumps the version to 0.7.27 and incorporates various bugfixes.
0.7.261 fixThis release primarily focuses on an asynchronous fix for batch sandbox creation.
0.7.25This release bumps the version number from 0.7.24 to 0.7.25.
0.7.241 fixThis minor release primarily addresses a bug fix related to database initialization.
0.7.23This release bumps the version number from 0.7.22 to 0.7.23 as a routine maintenance chore.
0.7.22This release bumps the version number from 0.7.21 to 0.7.22.
0.7.211 featureVersion 0.7.21 introduces a minor enhancement by passing the MCP environment variable to the STDIO MCP client.
0.7.201 featureThis release introduces Node.js support for node-based MCPs. It primarily focuses on adding new functionality rather than fixing existing issues or introducing breaking changes.
0.7.19This release primarily involves bumping the version number to 0.7.19.
0.7.181 fixVersion 0.7.18 addresses an asynchronous size bug and bumps the version number.
0.7.17This release primarily consists of maintenance chores, including updating the version to v0.7.17 and adjusting CI configurations for integration tests.
0.7.162 fixesVersion 0.7.16 primarily addresses two specific bugs concerning Anthropic streaming and SleepTimeMultiAgent tool usage.
0.7.151 fixVersion 0.7.15 primarily addresses a bug in the Docker Compose configuration related to the external database setup.
0.7.14This release (0.7.14) is a maintenance chore release, likely containing minor internal updates or dependency bumps.
0.7.13This release (0.7.13) is a maintenance chore release, likely containing minor internal updates or dependency bumps.
0.7.12This release primarily bumps the version number to 0.7.12.
0.7.11This release bumps the version to v0.7.11.
0.7.102 fixesThis is a bugfix release addressing issues found in Gemini Vertex integration and summarization capabilities.
0.7.9This release primarily bumps the version number to 0.7.9.
0.7.81 fixThis release primarily contains a bug fix ensuring in-context messages are correctly trimmed to the cutoff. The version has been bumped to 0.7.8.
0.7.7No release notes provided.
0.7.62 fixes1 featureThis release includes fixes for Anthropic streaming and Gemini model retries, along with new support for custom tool execution environments.
0.7.52 fixes1 featureThis release introduces new API endpoints for resource counting and addresses critical bugs related to agent deletion race conditions and Azure context window handling.
0.7.42 fixes1 featureThis release focuses on stability and correctness, fixing issues related to sleeptime input locking and LLM payload parameter passing, alongside an enhancement for local tool execution environments.
0.7.3This release (0.7.3) is a maintenance chore release, likely containing minor internal updates or dependency bumps.
0.7.2This release primarily consists of version bumps (0.7.1 to 0.7.2) managed through chore updates.
0.7.11 fix2 featuresThis release introduces support for reasoning tokens with Gemini Flash and fixes an issue with message filtering in sqlite3 conversation searches. It also includes the addition of a database compose file.
0.7.0This release primarily consists of maintenance chores, including updating the issue template and releasing version 0.7.0.
0.6.54This release (0.6.54) is a maintenance chore update, likely containing minor fixes or internal changes.
0.6.53No release notes provided.
0.6.52This release primarily consists of documentation updates and a version bump to 0.6.52.
0.6.51This release (0.6.51) is a maintenance chore release, likely containing minor internal updates or dependency bumps.
0.6.502 fixesThis release includes a fix for setting sequence IDs in sqlite3 using event listeners and addresses issues found in summarization for Google models.
0.6.49This release (0.6.49) is a maintenance chore release, likely containing minor internal updates or dependency bumps.
0.6.481 fixThis release includes a bug fix addressing incorrect parsing of system messages when using the Anthropic provider.
0.6.47This release (0.6.47) appears to be a maintenance or chore release, likely containing minor internal updates or dependency bumps.
0.6.461 featureVersion 0.6.46 introduces an enhancement to conversation search functionality to include agent messages. This release also marks the first contribution from a new community member.
0.6.45This release (0.6.45) is a maintenance chore release, likely containing minor updates or dependency bumps.
0.6.44Breaking7 fixes5 featuresThis release introduces logging for bedrock, updates to use AzureOpenAI for listings/completions, and addresses several API inconsistencies and bug fixes, including respecting configured embedding models.
0.6.432 fixesVersion 0.6.43 primarily focuses on bug fixes, including ensuring the parts list is never empty in a dictionary conversion and cleaning up optional dependencies.
0.6.421 fix1 featureThis release focuses on improving stability by fixing uvicorn worker issues and enhancing the logging system.
0.6.411 featureThis release integrates the OpenTelemetry collector directly into the letta image for enhanced observability.
0.6.402 featuresThis release includes fixes for archival statistics and Uvicorn environment variables, alongside updates to testing infrastructure.
0.6.393 fixes1 featureThis release introduces MCP support and includes several bug fixes, notably addressing issues with directory loading and Claude empty chunk warnings.
0.6.371 fix1 featureThis release includes various fixes and removes the length limitation for agent names.
0.6.36This release bumps the version to 0.6.36.
0.6.35This release bumps the version number from 0.6.34 to 0.6.35.
0.6.341 featureThis minor release bumps the version and increases the maximum token limit available in the system.
0.6.332 fixes2 featuresThis release introduces new type resolution capabilities and partial support for Claude 3 Sonnet, alongside fixes for Pydantic schema generation and argument handling.
0.6.32This release primarily consists of dependency updates and a version bump to 0.6.31.
0.6.31No release notes provided.
0.6.302 fixesThis release bumps the version to 0.6.30 and includes fixes related to making OpenTelemetry dependencies required and ensuring SQLite compatibility by changing jsonb to json.
0.6.291 featureVersion 0.6.29 introduces user identity features and updates the header for identity retrieval requests.
0.6.281 fix2 featuresThis release introduces support for the gpt-4o-mini context length mapping and integrates new providers, bedrock and deepseek, alongside various bug fixes.
0.6.271 fix1 featureThis release fixes an issue with VLLM usage and enhances the reliability of Vertex AI models by setting the connection mode to ANY.
0.6.262 featuresThis release introduces an example notebook for tool rules and applies a patch to the Google Vertex integration, bumping the overall version.
0.6.251 featureThis release introduces support for Vertex AI integration. The full details are available in the comparison link.
0.6.243 fixes1 featureThis release focuses on bug fixes across error handling, Google embeddings, and CLI operations, alongside improvements to the Anthropic integration.
Common Errors
ModuleNotFoundError6 reportsThe "ModuleNotFoundError" in MemGPT typically indicates a missing Python package required by the application. To fix it, identify the missing module (e.g., asyncpg, mcp) and install it using pip: `pip install <missing_module>`. Ensure you're installing packages within the correct environment (e.g., venv) to avoid conflicts.
BadRequestError3 reportsBadRequestError in memgpt usually arises from malformed API requests to the LLM provider, such as exceeding token limits, incorrect formatting of the request body, or providing invalid arguments to a function call. To fix this, carefully inspect the request being sent to the LLM, ensure it adheres to the provider's API specifications (including token limits and data types), and validate that tool call arguments are correctly formatted and complete before submission. If using a proxy, confirm it's correctly configured and forwarding requests without modifications that cause errors.
NameError2 reportsNameError usually arises when a variable or function is used without being defined in the current scope. To fix this, ensure the variable or function is defined (imported or declared) before its usage. Alternatively, check for typos or incorrect capitalization in the variable or function name and correct them to match the definition.
HandleNotFoundError2 reportsThe "HandleNotFoundError" in MemGPT often arises when the requested resource (like a specific LLM model) isn't available in the current provider's supported options or hasn't been correctly configured. To resolve this, verify that the model name in your `config.json` or agent creation parameters is supported by your chosen provider (e.g., OpenAI, Azure), and that you've properly configured required API keys or authentication details for that provider. Update your configuration with a valid model name and working credentials, or switch to a provider that supports your desired model.
AttributeError2 reportsAttributeError usually indicates that you're trying to access a non-existent attribute or method of an object. This often happens due to typos in attribute names, incorrect object types, or outdated library versions. To fix it, double-check the attribute name and object type, ensure the library is up to date, and verify the attribute exists in the object's class definition or parent classes.
UniqueViolationError2 reportsUniqueViolationError usually arises when attempting to insert data with a primary key or unique constraint that already exists in the database. To fix it, either update the existing record instead of inserting a new one if the data represents an update, or ensure your application logic prevents duplicate insertions by checking for existing records before creating new ones. Consider using `upsert` functionality provided by your database ORM if updating behavior is intended; otherwise, implement a `get_or_create` or similar pattern.
Related AI & LLMs Packages
AutoGPT is the vision of accessible AI for everyone, to use and to build on. Our mission is to provide the tools, so that you can focus on what matters.
Get up and running with OpenAI gpt-oss, DeepSeek-R1, Gemma 3 and other models.
🦜🔗 The platform for reliable agents.
The most powerful and modular diffusion model GUI, api and backend with a graph/nodes interface.
LLM inference in C/C++
GPT4All: Run Local LLMs on Any Device. Open-source and available for commercial use.
Subscribe to Updates
Get notified when new versions are released