Deleted agent memories keep resurfacing via derived summaries, dev warns
govinargo · reddit · 2026-11-19
A developer details a production-grade pain point: wrong facts asked of AI agents persist even after deletion, because summaries and derived memories built on them stay behind, with cleanup quality varying by model (Opus better than Haiku).
- Attempted fixes fell short: MCP + SQLite app layer had poor performance on long hierarchies; Neo4j/Cypher leaves orphaned memories or blocks deletes; Mem0's source tags only catch direct facts — open GitHub issues exist on both Mem0 and Graphiti
- Opus 5.5 can at least warn about suspected memory poisoning, but cleanup misses cases in long synthetic conversations
- Conclusion: agent memory engines need revoke, integrity, and agent isolation built in from day one — the author is now writing a small storage engine designed around exactly that
More from coding & agent
- rakyll: the biggest coding skill now is knowing where to look — rakyll · 2026-11-19
- Illustrated Guide to Agent Skills hits English release after 100k Chinese copies — dotey · 2026-11-19
- Anthropic, Garry Tan and Addy Osmani publish their full Claude Code configs — 7 free repos worth forking — garrytan · 2026-11-19
- The Resolution Hypothesis: agent evolution stalls because noisy evals can't spot real improvements — ShanghaiTech · 2026-11-19
- Microsoft ships Copilot SDK 2.0 with Rust core, powering all Copilot agents — PaulShellDev · 2026-11-19
- Rediscovering Notion as a Full Context Front End for AI Workflows — ivanhzhao · 2026-11-19