LLM Agent Memory

LLM agent memory is the broader design problem of how agents store, retrieve, update, and compress context over time without drowning themselves in their own notes. The first Mercury source frames markdown-heavy memory as acceptable for humans but increasingly wasteful for agents at larger scale.

This page should eventually compare wiki-style memory, vector retrieval, structured stores, event logs, and hybrid systems. Right now it mainly exists to anchor the Mercury Memory System page in a bigger map and connect it to GitHub Projects Community as the initial source.

2026-04-28 (from raw/2026-04-28-github-projects-mercury-memory-system.md): The Mercury pitch contrasts agent memory needs with Karpathy-style second-brain workflows.