Don't applaud cross-session memory yet: compaction drops decision boundaries
sujingshen · x · 2026-09-16
The author proposes a colder split for memory upgrades like cross-window notes, context compaction, and summary chains:
- Summary memory remembers roughly what happened.
- Decision memory remembers why you did it that way, especially why you didn't.
What compaction most easily loses is the latter: rejection reasons, aesthetic red lines, constraints. Once lost, the agent "gets the gist" but doesn't get you, and reinvents pitfalls you already hit.
So when you see "cross-session memory," don't applaud first; ask whether what survives compaction is a story summary or your decision boundary. The author asks harness/agent builders to share real cases of key constraints lost after compaction.
The quoted post notes Codex harness will roll out advanced context management in coming weeks, keeping notes across context windows without repeatedly compressing into a single summary, and wonders whether this is just making compaction a linked list of summaries or adding more prompting around file-based / MemGPT-style memory.
Related event: Structured key-value memory beats prose for agents(2 posts)→
More from AGI Musings
- David Patterson argues AI and robots will boost economic output 100x in 20 years, funding universal high income — davidpattersonx · 2026-09-16
- Gary Bernhardt: cognitive complexity sets the ceiling, for humans and AI alike — sidjustice_ · 2026-09-16
- Predicting AGI by tracking top lab alumni migration patterns — dosco · 2026-09-16
- Mathematician satirizes AI benchmarks: 'we never solved problems for benchmarks' — RexDouglass · 2026-09-16
- Domingos: The biggest AI catastrophe skeptics are cybersecurity and bio experts — pmddomingos · 2026-09-16
- Guardian cartoon skewers techbro doom-mongering: AI apocalypse talk as marketing — nordicinst · 2026-09-16