Hindsight v0.10.2: Long-term memory that lets your agent actually learnHindsight v0.10.2：给智能体一套会「学」的长期记忆

Ordinary vector libraries only do "look for similar sentences". Hindsight breaks down memory into three layers: world facts (what the world is like), empirical facts (what the agent has done itself), observations (the judgments it derives from them), and then overlays an entity relationship diagram to connect people, projects, and events.

The retrieval uses TEMPR multi-strategy fusion: semantic retrieval, keyword retrieval, graph retrieval, and timing retrieval. The four results are sorted by RRF merging, so the time-bound question "What happened last spring?" can also be hit.

There is also a rare ability called disposition traits: the degree of suspicion, literalism, and empathy can be configured, which directly affects the way it forms opinions. Writing assistants and reviewers need two different personality settings.

Docker is the most convenient: docker run -it --name hindsight --restart unless-stopped -p 8888: 8888 -p 9999: 9999 -v hindsight-data:/home/hindsight/.pg0 ghcr.io/vectorize-io/hindsight: latest, service from 8888, built-in PostgreSQL, data falls on the data volume.

To install hindsight-api into an existing Python environment, pip install hindsight-api, and set hindsight_API_LLM_provider and hindsight_API_LLM_API_key to run hindsight-api directly. The LLM provider supports openai, anthropic, gemini, groq, ollama, lmstudio, github-copilot, and the model defaults to gpt-4o-mini.

MCP is already built-in: the service endpoint http://localhost: 8888/mcp/{bank_id}/, which directly exposes the three actions of retain, recall, and reflect as tools for any MCP client; if you don't want to think of the full service, there is also the hindsight-local-mcp version of stdio.

The diagram is a schematic diagram of the official memory architecture (retain/recall/reflect three links).

The official command for one-click access to programming agents is provided: npx @ vectorize-io/hindsight-coding-agents install all installed agents and connect them one by one; use npx @ vectorize-io/hindsight-coding-agents install claude-code only to connect Claude Code.

After receiving it, the preferences, project conventions, and pits that you talked about in the previous session will be automatically brought back to the next session - the biggest difference between it and "explain it from zero every time you open a new window".

The download button provides the official PyPI sdist package hindsight_api_slim 0.10.2 (about 1.77 MB), corresponding to v0.10.2 released in release 2026-09-29.

Local installation: pip install./hindsight-api-slim-0.10.2.tar.gz. The slim version removes the bulky optional dependencies and is suitable for self-built servers; just write the client and install hindsight-client.

there are also separate executables for Linux and macOS in the release, but Windows binaries are not officially available. Windows users should take the Python or Docker route.

This article is compiled from public sources and ships with the matching resource. Product features and pricing are subject to the official page. Resources are for learning and exchange only — please respect the original license.