RFC: Hindsight long-term memory for Flue agents

[benfrank241]

Summary

Add optional long-term memory for Flue agents backed by Hindsight, an open-source agent memory system, so agents can recall durable facts across sessions rather than only within a sandbox run.

Background & Motivation

Flue is a sandbox agent framework — great for scoped, ephemeral runs. But agents often benefit from memory that persists across runs: durable facts about a user/domain and consolidated knowledge, recalled on demand. Today a Flue user would wire that up by hand.

Hindsight organizes memory biomimetically into world facts, experience facts, and consolidated mental models, with multi-strategy recall (semantic + BM25 + graph + temporal) and a reflect operation. It exposes a REST API and an MCP server, so it slots in without Flue taking on a storage dependency.

Goals

• Let a Flue agent recall relevant long-term memories into context before a model call.
• Let a Flue agent retain new memories after a run (entity extraction + consolidation handled server-side by Hindsight).
• Keep it fully optional and side-effect-free when not configured.

Example

Illustrative only — exact surface to match Flue's conventions:

import { withHindsight } from "@hindsight/flue"; // illustrative

const agent = withHindsight(baseAgent, {
  apiKey: process.env.HINDSIGHT_API_KEY,
  bank: "user-123", // isolated memory store per user/agent
});

Prior art

Hindsight already ships as a first-party memory provider in NousResearch's Hermes agent, alongside honcho, mem0, supermemory, and byterover. I'm the maintainer; opening this as an RFC per CONTRIBUTING to gauge interest before proposing anything concrete.