Scanning 23 agent sessions, a dev found 3 silent failure modes in memory systems
No_Advertising2536 · reddit · 2026-09-11
A developer scanned 23 of his own Claude Code sessions and found four recurring workflows — "debug from logs" was re-taught from zero in 16 of them, because rules files only capture the stack, not how the work goes here. He built a memory layer that extracts these sequences and shows the agent past success rates before matching shell commands.
Three silent bugs emerged:
- Failures never fired the recorder. The host only sends the post-tool event after successful commands (verified with a marker command exiting 7), so track records built from events can only climb and a broken workflow stays invisible.
- A single shared word broke the confirmation gate. Replaying 3,456 real commands, naive matching would have interrupted 795 (1 in 6); requiring the command to actually be a step cut it to 8.
- Retrieval couldn't return nothing, so short prompts like "so now?" always pulled three unrelated entities.
His takeaway: all three failed silently, and two were reported as healthy by his own status checks. He also notes every memory tool he's seen stores steps but never outcomes.
More from coding & agent
- As Claude Code and Codex get scary good, what's left for small teams building agent products? — oliver_dev · 2026-09-11
- Dev builds talk on guardrails workflow for shipping AI-written code without reading it — TejasKumar_ · 2026-09-11
- banteg: Codex auto-review has regressed, blocking steps needed to complete authorized tasks — banteg · 2026-09-11
- A doc-anchored agent workflow: you write, the agent only critiques and finds disagreements — lucasmeijer · 2026-09-11
- AI sales agents shine at the top of funnel but lose real deals, says GTM practitioner — gogeta7124 · 2026-09-11
- SymKit MCP: 44 tools for AI agents to verify symbolic derivations — Foreign-Specific-604 · 2026-09-11