Repository navigation
Replies: 2 comments
|
Implementation of this proposal: #6406 (native memory provider, transcript bridge, owner-safe maintenance, and |
|
+1 for official Hermes support, from someone running it in production. Two parts of this proposal have landed: the harness registry entry (#5292) and transcript ingest on writer-claimed brains (#5235, v0.60.163.0). The ingest filters did not land. #5119 (filter by session source) and #5118 (turn-level cutoff) were closed in the issue census and moved to Ideas (#8175, #8174). Master's Our setup: a gbrain 0.60.728 host running Hermes v0.23.1. It has several Hermes profiles, some under separate OS users as thin clients. What we still have to work around:
I understand why #6406 was more surface than the project wants to maintain. Items 1–3 are small and harden code gbrain already ships. I'm happy to test. I can run release candidates against several live Hermes stores and report back. I can also check a pinned schema against a current Hermes release and share schema-only output, with no message content. |
Uh oh!
There was an error while loading. Please reload this page.
Moved from #6065, opened by @barronlroth on 2026-10-05. Feature requests now live in Discussions > Ideas.
Problem
Running GBrain under Hermes Agent still means maintaining a downstream stack of glue code that breaks on most upgrades. The pieces are individually tracked, but there is no single supported integration target, so each Hermes user rebuilds the same scaffolding:
docs/mcp/HERMES.md) but absent fromsrc/core/harness/registry.ts, sogbrain mcp adaptersand every registry-driven surface ignore it (Hermes is documented (docs/mcp/HERMES.md) but absent from the harness registry (src/core/harness/registry.ts) #5292).~/.hermes/state.db). Native ingest → Dream was fixed in v0.60.36.0 ([dream] Native transcript ingestion never reaches Dream, silently disabling its core conversation-to-wiki synthesis #4419), but ingest still aborts on writer-claimed brains (gbrain transcripts ingest aborts on writer-claimed brains: the importer still writes through the legacy writer (writer_coordinator_required) #5235), cannot filter by session source (Hermes transcript ingestion cannot filter sessions by source #5119), and has no turn-level cutoff (Support a turn-level cutoff for Hermes transcript ingestion #5118). In practice users keep custom exporters intosession_corpus_dir.facts-absorbjobs that no daemon drains in a one-shot PGLite setup, and dream wrappers that guard on nonterminal jobs then refuse to run. Every user writes their own lock reconciliation, drain, and retry wrapper..gbrain-owner.json) without an activated writer, which then refuses lint/sync/facts withwriter_coordinator_required. A Hermes-hosted single-user brain needs a clear, documented "stay unmanaged" or "activate" path.~/.hermes/auth.json→credential_pool.openai-codex). There is no upstream provider recipe, so this lives in forks.Proposal
A first-party Hermes integration (a
gbrain-hermesplugin or a documented, CI-covered integration target) that provides:state.dbinto a dream-eligible corpus. It should use stable IDs and a cursor, filter by source and cutoff (Support a turn-level cutoff for Hermes transcript ingestion #5118, Hermes transcript ingestion cannot filter sessions by source #5119), work on writer-claimed brains (gbrain transcripts ingest aborts on writer-claimed brains: the importer still writes through the legacy writer (writer_coordinator_required) #5235), and fail loudly instead of producing a silent zero-input dream run.hermes-state-db-2026-08spec to a pinned Hermes schema target with a realstate.dbfixture in CI.openai-codexgateway recipe reading Hermes-managed credentials, priced at $0 marginal cost in budget metering. I know earlier PRs (feat(ai): add OpenAI Codex OAuth gateway recipe #977, dream: support Codex OAuth synthesize on PGLite #2214, feat(ai): codex-cli recipe — GPT chat via Codex CLI OAuth subscription #3587, feat: add OpenAI Codex OAuth provider #698, feat(ai): openai-codex — ChatGPT-plan OAuth provider for chat + expansion #1643) were closed. I'm listing it for completeness; the rest of the proposal stands without it.Why upstream
These are generic Hermes-host needs, not personal configuration. Today they exist as patch stacks that need a re-audit on every release. I run this stack daily (PGLite, nightly dream, Codex OAuth) and am happy to contribute the pieces I've built and tested, with regression tests.
Related
#5292, #5235, #5119, #5118, #4419, #4551, #4149, #2213
All reactions