Dev ships three Hermes plugins to stop agents from writing false memories and loop garbage
yleven120 · reddit · 2026-10-21
The author runs an AI agent daily that writes to its own memory and skills, and identified three recurring failure modes — then built one Apache-2.0 plugin for each and sent them upstream, with two merged into the official catalog:
- hermes-degeneracy-guard: sits at pretoolcall, scores tool-call arguments, and routes degenerate loops to Hermes' own approval gate instead of guessing; cron runs fail closed by default, preventing corrupted file writes.
- hermes-provenance-gate: blocks memory/skill writes like "the user decided X" that the user never made, requiring a verbatim quote from the user's actual session input. On 133 labeled lines: precision 1.000 (34/34), recall 0.694 (34/49); misses are published alongside regression tests.
- hermes-skill-lineage: records per-write content fingerprints, sources, feature links, and a seq→prevhash→hash chain; trace/descendants/report/verify/head commands answer where a skill version's text came from and whether the chain is intact. 48 tests pass; end-to-end logs committed.
The author also reports that Hermes' repetition-loop stream guard is not wired into the Bedrock Converse and Codex Responses streaming paths, with a fix PR open.
More from coding & agent
- Index Network open-sources Social Discovery Protocol: agents negotiate fit on your behalf — carsonfarmer · 2026-10-23
- Installed a formal-verification skill: it silently installed Java, MacBook runs 10 degrees hotter — zealcaiden · 2026-10-23
- OpenAI's Decisions API hits GA: near real-time routing up to 10x faster than GPT-6 Luna — OpenAIDevs · 2026-10-23
- Ruff author warns: upgrade Node/Bun now — --use-system-ca still accepts the eMudhra cert — charliermarsh · 2026-10-23
- Proof of Agent: paste one prompt and your AI mines, grades, and self-governs — adamamcbride · 2026-10-23
- DeepLearning.AI weekly: GPT-6.1 hits 97.9% byte parity on COBOL migration — DeepLearningAI · 2026-10-23