news.inema.pro / All news

project Published on 15 October 2026

hindsight: local memory for your agent to learn through Claude use

hindsight gives an agent memory to learn through use, locally and with a Claude subscription.

Hindsight is a long-term memory system for AI agents, created by Vectorize and released under the MIT license. It stores experiences, searches for information, and reflects on what has been recorded. Over time, it consolidates repeated facts into observations and evidence-backed beliefs. The memory stays on the user's machine; a language model is part of the process.

For people building agents, bots, or assistants, this can help carry conversations and decisions across sessions. Instead of relying only on chat history, an agent can retrieve information by meaning, keywords, links between entities, or time period, then use what it has learned in new responses. Each user, customer, agent, or project can have a separate memory database. The system can also block personal data before storing it.

To get started, you need Python 3.11 or Docker and a language model. The source describes a local setup using a Claude subscription, without an API key, and also mentions a Codex subscription or a local model. For Portuguese and Spanish, the documentation recommends configuring the BAAI/bge-m3 embedding model. Before connecting it everywhere, note that each memory entry calls the model and uses some of the subscription quota; the documentation describes the claude-code provider as intended for personal, local use.

Read at the source →

The link opens the original content; some require a subscription.