5-week A/B test: OptMem vs Claude Code native memory — slightly better, rotting differently
_rchaves_ · x · 2026-09-14
A LangWatch developer replaced Claude Code's native memory with @VictorTaelin's OptMem for 5 weeks, and leveraged sessions where OptMem randomly failed to activate to run a natural A/B test against 3 months of prior native-memory data.
Key findings:
- OptMem beats native memory, but not by much, and in different ways
- Memory helps agents mainly in long sessions
- Biggest value: avoiding rediscoveries and time-wasting "traps"
- Token cost savings are modest
- The author is sticking with OptMem
OptMem's core design keeps memory at a fixed size: structured and human-readable, unlike free-form markdown or compressed vector DBs, which caps the blast radius of memory rot. Quoting Claude's own analysis: native memory corrects via in-place rewrites (1,719 of 2,081 writes were rewrites; MEMORY.md rewritten 601 times), but its rot is silent staleness as the world moves (39% of path tokens went stale; 68 of 260 files fell out of the index). OptMem is append-only, so its rot shows up as contradictions and compression (only 3 explicit corrections in 281 lines, 5 unmarked contradicting pairs). Curiously, OptMem still accumulated fewer misleading memories.
More from coding & agent
- The key coding skill now: directing agents well — and knowing when not to trust them — YvesMulkers · 2026-09-14
- chrome-mcp fills a 30-field form in real Chrome: 148 logged, auditable actions — SYEDNIICK · 2026-09-14
- Building a Local Router With Smart Fallbacks to Stop Rate Limits Killing Your AI Agents — phanue1 · 2026-09-14
- A SQLite/S3 filesystem with per-directory OpenPolicyAgent scripts for AI agents — tobowers · 2026-09-14
- Processing SIEM tickets with AI: workflow tool or pure agent? — Melodic_Tooth6885 · 2026-09-14
- Same 3D simulation prompt: Agnes 2.5 Pro Beta costs ~$0.20 vs ~$1.70 on GPT-5.6 Sol — iamaliveix · 2026-09-14