Tutorial: Give your agent long-term memory in four steps (self-hosted Hindsight)教程：四步给智能体接上长期记忆（自建 Hindsight）

First make sure Docker is available and then execute: docker run -it --name hindsight --restart unless-stopped -p 8888: 8888 -p 9999: 9999 -e hindsight_API_LLM_API_key = $ OPENAI_API_key -v hindsight-data:/home/hindsight/.pg0 ghcr.io/vectorize-io/hindsight: latest.

8888 is the rest and MCP port, 9999 is the console port, the data volume .pg0 hangs the built-in PostgreSQL, and the container restart memory will not be lost. --restart unless-stopped Guarantee boot.

I don't want to use Docker to pip install hindsight-api in the Python environment, set hindsight_API_LLM_provider and key to run hindsight-api, and listen to 8888 by default. The picture is a schematic diagram of its memory link.

Having a separate memory bank for each use is the easiest thing to do (one assistant, one library). When building the library, fill in the name and a background description, this background will participate in the follow-up judgment as its self-perception.

Write with retain, throw in the fact that "the user prefers to use Python for data science projects"; query with recall, ask "what programming language the user prefers" to see if it can be retrieved; use reflect when it is needed, such as "This machine learning project should recommend Python or R" - it will give a tendency judgment based on accumulated memory, rather than paraphrasing the original text.

You can try three lines of code on the Python side: from hindsight_api import MemoryEngine, await memory.initialize (), and then call create_memory_bank/retain/recall in turn.

The MCP address that comes with the service is http://localhost: 8888/mcp/{bank_id}/. Replace {bank_id} with the ID of the library you just built, and retain, recall, and reflect will become three callable tools directly. Any MCP-compatible client can hang up.

Programming agents have a more convenient connection method: npx @ vectorize-io/hindsight-coding-agents install claude-code is only connected to Claude Code, and install all automatically detects that all the agents of the machine are connected one by one.

If you don't want to reside in the service, the official also offers the stdio version of hindsight-local-mcp, which is suitable for temporarily pulling up the memory ability only when needed.

The verification method is very straightforward: say a preference in session A (for example, "This project uses pnpm, do not use npm"), close the session, open a new session B and ask "What package manager does this project use" directly, if you can answer correctly, it means that the memory link is connected.

The second step is to change the model: change the hindsight_API_LLM_provider to ollama or lmstudio, and run it with the local model to be completely offline - memory and reasoning are on your own machine, and sensitive items are especially suitable.

Last reminder: The memory system will write the conversation into the database and send it to the model you configured. Please confirm the compliance boundaries before accessing the company's internal data.