What area does this relate to?
An existing contribution (recipe, schema, dashboard, integration, skill)
What would you like to see?
Title: Feature Request: Hindsight as an alternative memory backend?
Body:
Hey folks,
I've been looking at Open Brain as a Life OS setup for personal use, and while going through the architecture I started wondering about the memory retrieval side of things.
Right now OB1 relies on pgvector for semantic search, which works — but I've also been using Hindsight (https://github.com/vectorize-io/hindsight) for agent memory in coding workflows, and it takes a pretty different approach: four parallel retrieval strategies (semantic, BM25, graph traversal, temporal), cross-encoder reranking, and a reflect operation that actively consolidates knowledge into observations and mental models.
This got me thinking:
- Has anyone explored using Hindsight as a memory backend for Open Brain instead of (or alongside) raw pgvector?
- Would OB1's extensions and capture pipeline work with a different retrieval engine under the hood?
- For people who already run Hindsight — how do you avoid the double-token problem when both systems are connected via MCP?
I'm genuinely curious whether swapping the retrieval layer could make OB1 stronger without losing what makes it great (the import recipes, the extensions ecosystem, the capture flow). Or whether there are good reasons pgvector is the right choice here that I'm not seeing.
Would love to hear thoughts from anyone who's tried both, or from the maintainers on whether this kind of pluggable backend is even on the radar.
Alternatives considered
Continuing with pgvector as-is and accepting the retrieval limitations. Or running both Hindsight and Open Brain side by side via MCP — but that creates a double-token problem where every conversation pulls context from both systems redundantly.
What area does this relate to?
An existing contribution (recipe, schema, dashboard, integration, skill)
What would you like to see?
Title: Feature Request: Hindsight as an alternative memory backend?
Body:
Hey folks,
I've been looking at Open Brain as a Life OS setup for personal use, and while going through the architecture I started wondering about the memory retrieval side of things.
Right now OB1 relies on pgvector for semantic search, which works — but I've also been using Hindsight (https://github.com/vectorize-io/hindsight) for agent memory in coding workflows, and it takes a pretty different approach: four parallel retrieval strategies (semantic, BM25, graph traversal, temporal), cross-encoder reranking, and a reflect operation that actively consolidates knowledge into observations and mental models.
This got me thinking:
I'm genuinely curious whether swapping the retrieval layer could make OB1 stronger without losing what makes it great (the import recipes, the extensions ecosystem, the capture flow). Or whether there are good reasons pgvector is the right choice here that I'm not seeing.
Would love to hear thoughts from anyone who's tried both, or from the maintainers on whether this kind of pluggable backend is even on the radar.
Alternatives considered
Continuing with pgvector as-is and accepting the retrieval limitations. Or running both Hindsight and Open Brain side by side via MCP — but that creates a double-token problem where every conversation pulls context from both systems redundantly.