Topic
An approach to designing memory systems for LLM agents, separating short-term in-task state from long-term cross-session recall and defining retrieval rules.
The estimate was 30,000 to 40,000 credits per simulation, the gateway needed only three read patterns, and the store already holding tenant rows could serve all three, so tenancy became an argument every query passes.
Reality
What OpenCode users call memory is an instruction loader, a SQLite event log and a compaction summarizer. Each has a different owner and a different way of losing your project rule.
Reality
A principal AI researcher says his team swapped a graph for a tree-structured memory to get faster incremental updates. That makes agent memory a write problem, not a retrieval contest.
Reality
A dev.to writeup drops vector stores for a SQLite table with tag and timestamp columns. Its own numbers put the practical ceiling at a million rows, not a hundred million.
Reality
A dev.to post argues coding agents fail because abandoned decisions and standing conventions come back with identical confidence. The proposed fix is a schema with explicit precedence.
Reality
A dev.to writeup argues the expensive agent bug is ambiguity after a timeout, and that durable per-action outcome records, not larger memory, are what stop duplicate writes.
Reality
A dev.to series on harness engineering splits persistence into in-task scratch and cross-session recall, and argues that building them as one system produces neither.
Reality