Hindsight

What vectorize-io/hindsight MCP server does

The vectorize-io/hindsight MCP server is built around persistent memory for AI agents. It is intended for agents that need to learn from information over time, not simply replay previous conversation turns. The repository describes three central operations: retain, recall, and reflect.

Retain stores information in a named memory bank. Recall searches that bank for relevant information, while reflect generates a response based on the stored memories and the request. The README demonstrates retaining facts about a person, asking a question about those facts, and generating a natural-language response from the same bank.

The project also documents memory types, observations, mental models, knowledge pages, and memory banks. These concepts support organizing information and using it in later agent interactions, although the provided material does not enumerate the individual MCP tool definitions.

How it works

Hindsight can run as a service with an HTTP API. A Python or Node.js client connects to the service using its base URL, then identifies a memory bank when calling retain, recall, or reflect. The default self-hosted example uses http://localhost:8888 for the API and exposes a separate interface on port 9999.

The service uses an LLM provider for its memory-related processing. Documented provider options include hosted services such as OpenAI, Anthropic, Gemini, Groq, Bedrock, and Vertex AI; local runtimes such as Ollama, LM Studio, and Llama.cpp; OpenAI-compatible endpoints; and gateways including LiteLLM. Existing subscriptions can also be used with the documented provider names for OpenAI Codex, Claude Code, and GitHub Copilot.

Setup and configuration

The vectorize-io/hindsight MCP server can be deployed with the published Docker image, installed with pip install hindsight-api, or deployed through the provided Helm chart. The Docker example persists data in a volume and maps API port 8888 and UI port 9999. An external PostgreSQL deployment is also documented, and Oracle AI Database is listed as an enterprise storage option.

The standard examples configure HINDSIGHT_API_LLM_API_KEY. Provider selection uses HINDSIGHT_API_LLM_PROVIDER. A hosted Hindsight Cloud option is available at https://api.hindsight.vectorize.io, where clients connect with an API key instead of deploying the service themselves.

Python users can install hindsight-all for an embedded mode that does not require a separate server. Intel Macs require the documented hindsight-all-slim package.

Tools and capabilities

• Store information in named memory banks with retain.
• Search stored memories with recall.
• Generate disposition-aware responses with reflect.
• Connect through Python, Node.js/TypeScript, Go, CLI, or REST clients.
• Run with embedded storage, external PostgreSQL, or documented enterprise storage.
• Use hosted or local LLM providers, depending on deployment configuration.

Limitations and notes

The provided material does not list the MCP server’s exact tool names, transport configuration, or client-specific setup instructions. It also does not establish that every listed LLM provider works without its own credentials. Self-hosted examples require configuring an LLM provider, while local providers and some subscription-based provider modes may avoid a conventional API key.