Five places agent memory can live as loops replace prompt engineering
blaizedsouza · x · 2026-07-28
The idea
The thread argues that modern agent systems are no longer about writing a better prompt, but about designing loops and memory paths.
Five places agent memory can live
- Truncation — old turns fall out of the context window and are lost.
- Compaction — many turns are summarized back into the window.
- Retrieval — memory lives outside the window and is fetched by similarity.
- Files on disk — memory is stored by path and can be recovered exactly.
- Graphs — memory is stored as relationships rather than text.
Core point
The post emphasizes that every system must answer the same question: after a turn ends, where does the information go, and how do you get it back? It frames loops as the new center of agent engineering, replacing prompt-only thinking.
More from coding & agent
- Claude Code gets nicknamed the “grim grepper” — sull · 2026-07-28
- Developer jokes that Bible-style prompts might improve agent swarms — eigenhector · 2026-07-28
- Open-source JARVIS scaffold adds GitHub and Supabase-backed agent memory — johnbarber720 · 2026-07-28
- AI code migration ports geometry3Sharp to TypeScript and writes new tests — rms80 · 2026-07-28
- OpenAI Codex appears to be filling a sessions folder with 360 GB of data — HankYeomans · 2026-07-28
- Pricing pages now ship as Markdown so AI agents can parse them — stuffyokodraws · 2026-07-28