What do you think about using Hindsight additionally? What if I want to use structured memory with TEMPR search instead of local memory files?

[mkremnev]

I'll speculate)

[mnriem]

Moving this to discussion as it is an idea.

[dlukt]

I have never heard of either
Do you mean this?
https://github.com/vectorize-io/hindsight
I couldn't find anything for TEMPR

[mkremnev]

Yes, that's it. I use it everywhere now. What makes it special:

• a single window of knowledge in the form of graph relationships of your thoughts and developments
• fast semantic search across these graphs
• a unified context for your development approaches and development patterns. But you can also create unifications for individual projects

[dlukt]

What is TEMPR?
And you run 1 docker container for all? What do you use it with, I mean it requires API keys, so every request costs, or not? Couldn't this be adapted to use e.g. GPT-5-mini from copilot?

[mkremnev]

Multi-Strategy Retrieval (TEMPR) - memory extraction method

Main Strategies:

• Semantic Search: Vector similarity using cosine measure for shared meaning.
• Keyword Search: BM25 lexical search for exact term matching.
• Graph Search: Breadth-first traversal of relationships (entities, time, semantics, causality).
• Temporal Graph Search: Filtering by time intervals for queries with chronology.

I run Docker with a single command in unless mode, and the API is always available. Then, in the codebase, I use mcp, which accesses Hindsight via the API and calls the necessary functions (function calling).

Provider: openai, openai-codex, claude-code, anthropic, gemini, groq, ollama, lmstudio, vertexai

LLM models can also be changed.

[mkremnev]

They have very good documentation https://hindsight.vectorize.io/developer/configuration