Labsco
MCP SERVER

Hindsight

by vectorize-io ↗

Agent memory that learns rather than just recalls — retain, recall and reflect, exposed as MCP tools with one endpoint per memory bank.

Agent Memory
Summary
One MCP endpoint per bank is the isolation boundary, not a naming convention.

Because the endpoint carries the bank id, a client is wired to exactly one memory store and cannot reach another — which is what makes per-user or per-project memory safe to run on a shared server. Note that this is infrastructure you operate: it wants a server, a database and an LLM provider, and the local-model options exist precisely so memory contents need not leave your network.

What it is

A memory system for agents, with an MCP endpoint built into every server. Memories go in through retain, come back through recall, and reflect produces a considered answer over them. Facts do not stay a flat pile: they are consolidated in the background into observations that keep their supporting quotes and are refined rather than overwritten as evidence arrives.

What you get
  • Retain, recall and reflect available as tools to any MCP client, one endpoint per bank at `http://localhost:8888/mcp/{bank_id}/`, enabled by default
  • Banks as isolated memory stores — one brain per user, agent or project, with no cross-bank leakage, carrying background context and disposition traits that shape how reflect reasons
  • Observations: deduplicated beliefs built from many memories, each keeping exact supporting quotes and a proof count
  • Mental models and knowledge pages — standing answers to questions about a bank, rewritten in the background, read as a database lookup with no retrieval and no model call
  • Multilingual handling that preserves the input language and native scripts end to end
  • Memory Defense, an opt-in per-bank policy that scans every retain against 45 secret and PII patterns and either redacts or blocks the item before storage
Requirements

A running server plus an LLM provider key. The quickest path is Docker: run `ghcr.io/vectorize-io/hindsight:latest` with `HINDSIGHT_API_LLM_API_KEY` set, exposing the API on 8888 and the UI on 9999. `pip install hindsight-api`, Docker Compose with external PostgreSQL, and a Helm chart are the other options. `HINDSIGHT_API_LLM_PROVIDER` selects among 25+ providers including fully local ones like ollama and lmstudio; existing ChatGPT, Claude and GitHub Copilot subscriptions work without an API key. Hindsight Cloud is the managed alternative at `https://api.hindsight.vectorize.io`. Version 0.9.2.

Setup effort

One command plus a key — pip install hindsight-api, then supply credentials