Five silent agent-memory bugs from two months in production — "what's new" returned only last month's facts
ImaginationUnique684 · reddit · 2026-09-25
After merging six systems into one customer context layer, the author spent two months in pilot with real users and catalogued a set of bugs that never threw errors:
- "What's new" returned 10 facts, all from July: the graph held the last 30 days fine, but semantic+keyword search with no date term left ranking near noise — and 79% of the graph was July data. Fix: recency questions now get a wider pool sorted by each fact's own date, and only those.
- A cap of 40 kept the oldest end of the window, so the newest fact users saw was already 8 days old
- "A blocks B" and "finishing A unblocks B" stored as two separate facts
- A feature literally named Knowledge Graph extracted 0 facts per run, because the extractor thought it referred to its own memory
- A claim passed 3-of-3 review extraction, then died on the write's own extraction run
None threw an error. The author asks peers: what bit you first — time, duplicates, or extraction itself?
More from coding & agent
- Obsidian Starter Kit v4 turns your vault into a typed AI knowledge base with MCP and 375 skills — dSebastien · 2026-09-25
- RouterGrowth launches as 'OpenRouter for sales' — one agent prompt found 20 leads for $0.72 — aliscodes · 2026-09-25
- Dev says Claude remotely filed a lawsuit, answered emails and pitched his project with a demo video — aloncarmel · 2026-09-25
- Who's responsible when your AI agent spawns another agent? Multi-agent authority chains lack an audit trail — Fantastic-Sleep-3352 · 2026-09-25
- Developer says Codex has fully replaced Claude Code as his primary coding tool — eptwts · 2026-09-25
- Weco preprint: agent self-improves its own software 7 times over 8 days, gains transfer to 4 benchmarks — VraserX · 2026-09-25