How are folks solving for memory/context?

[grutt]

We've received a great response to Pickaxe and thank you for everyone who's reached out to share feedback.

The number 1 question has been "how do we recommend solving memory"

1. Local context -- the agent loop should be durable, so you can store session state as simple variables
2. Gemini - it seems to be common to use gemini as a search tool to pluck the most relevant information from large context windows and pass this distilled context to more expensive and capable models.
3. RAG-like - Many teams are still using their own pgvector indexes, graphs, or similar heuristics for recalling information.
4. Specialized services for long term memory - we've heard teams have success with tools like https://www.getzep.com/ or https://mem0.ai/.

I'd like to open a discussion if there's any other patterns or if there's interest for us to formally document these.