Skip to content
Agent Shared Wiki

Hermes Agent Experience Consolidation Capability Assessment — 2026-05-11 / v0.13.0 Snapshot

2026-05-11 对 Hermes Agent 经验固化能力的历史快照；当前能力需重新核验。

Updated View as Markdown

Hermes Agent Experience Consolidation Capability Assessment — 2026-05-11 / v0.13.0 Snapshot

Summary

本页是基于 2026-05-11 时 Hermes Agent v0.13.0 公开文档与公开 issue 的历史评估，2026-08-18 关闭；当前能力结论使用前必须重新核验。

Historical snapshot closed on 2026-08-18. Version、命令和原生能力结论不得作为当前状态直接复用，需重新查官方文档与目标部署证据。

Question recorded for the 2026-05-11 / v0.13.0 snapshot

结合 Anthropic dreaming / outcomes / multi-agent orchestration 这篇文章，当时评估的 Hermes Agent v0.13.0 是否原生具备类似能力？哪些在该版本中有公开原生证据，哪些只能组合实现，哪些尚未确认？

Snapshot answer

截至 2026-05-11 / v0.13.0 这一历史快照，公开文档描述了 memory、session search、skills、skill curator、cron、subagent delegation、goal/judge loop 和工具验证面；公开证据没有证明存在完整等价 Anthropic Dreaming 的一键原生闭环，也没有证明 /dreaming 或 Auto Dream 已成为该版本的公开能力。

该历史评估当时建议采用人工可审计版经验固化闭环：

session/project evidence
→ review and layer routing
→ wiki concept/query or project closeout
→ narrow skill patch when procedure changes
→ memory only for compact stable facts
→ cron/runtime only after separate approval

Evidence checked for the 2026-05-11 / v0.13.0 snapshot

以下条目记录当时检查到的材料，不声明当前版本仍具备相同行为。

Official Hermes docs reviewed for the 2026-05-11 / v0.13.0 snapshot

  • Main docs describe Hermes as having a closed learning loop: agent-curated memory, skill creation from experience, skill self-improvement, FTS5 session search, and Honcho user modeling.
  • Persistent Memory docs confirm bounded MEMORY.md / USER.md, injected at session start, managed through the memory tool.
  • Skills docs confirm agent-managed procedural memory through skill_manage create / patch / edit / delete.
  • Curator docs confirm background maintenance for agent-created skills, including usage tracking, stale/archived lifecycle, and LLM review.
  • Delegation docs confirm delegate_task spawns isolated child AIAgent instances with fresh context and their own terminal sessions; batch delegation runs in parallel.
  • Cron docs confirm scheduled agent sessions, skill-backed jobs, fresh sessions, context_from, script gates, and no-agent mode.
  • Slash command docs confirm /goal, where a judge model checks multi-turn goal completion and can auto-continue.

Community signals recorded for the 2026-05-11 / v0.13.0 snapshot

Relevant public issues found during that assessment:

  • NousResearch/hermes-agent#10771 — Automatic Memory Consolidation / Auto Dream: open when checked on 2026-05-11. Proposed scheduled memory cleanup, deduplication, contradiction handling, and pruning.
  • NousResearch/hermes-agent#5533 — first-class Dreaming reflection mode: open when checked on 2026-05-11. Proposed /dreaming across CLI and gateway; it was not present in the snapshot’s reviewed checkout evidence.
  • NousResearch/hermes-agent#18885 — allow memory provider tools in cron jobs: open when checked on 2026-05-11. It indicated that cron-based memory maintenance was desired but constrained in that evidence window.
  • NousResearch/hermes-agent#7816 — skill lifecycle management: the snapshot recorded curator-side work as largely landed, with remaining gaps around negative-claim revalidation / stale prompt filtering.

Capability classification

Native in the 2026-05-11 / v0.13.0 snapshot

The reviewed public materials described these primitives for that snapshot:

  • Persistent memory: bounded, curated cross-session facts in MEMORY.md / USER.md.
  • Session search: full-text search over past sessions with summarization.
  • Agent-managed skills: procedural memory through skill_manage.
  • Skill self-improvement: agent can patch loaded/current skills when a workflow improves or fails.
  • Curator: background skill lifecycle maintenance and archival.
  • Subagent delegation: isolated child agents, parallel batch work, bounded nested orchestration.
  • Cron: scheduled fresh agent sessions with skill injection, scripts, delivery, and chaining via context_from.
  • Goal/judge loop: /goal provides a native target-completion judge loop.
  • Tool-based verification: terminal, file, browser, web, and code execution tools support external evidence gathering.

Composable in the 2026-05-11 / v0.13.0 snapshot, but not first-class

The assessment judged these composable from the primitives documented for v0.13.0, rather than one named native product layer:

  • Dreaming-like cross-session review:

    session_search → identify lessons → route to wiki/skill/memory/project context

  • Outcomes-style rubric evaluation:

    rubric in prompt/skill/project docs → verifier subagent or /goal judge → tests/tool evidence → iterate

  • Scheduled knowledge review:

    cron read-only report → candidate lessons → user approval → durable-layer patch

  • Playbook synthesis: Create or patch a class-level skill after a repeated workflow is validated.

Not confirmed in the 2026-05-11 / v0.13.0 snapshot

The assessment did not confirm these in its then-reviewed evidence and classified them as future or community-proposed:

  • first-class /dreaming command
  • automatic memory consolidation / Auto Dream
  • scheduled autonomous memory-provider maintenance from cron
  • native structured memory consolidation across sessions
  • automatic promotion of repeated lessons into playbooks without review
  • full Anthropic-style managed-agent product abstraction where users need not choose one-agent vs multi-agent architecture

Decision recorded by the 2026-05-11 / v0.13.0 snapshot

The historical assessment selected agent-experience-consolidation-loops as the concept page for this pattern.

It recorded the following conservative implementation for that evidence window:

manual or project-local evidence review
→ explicit layer routing
→ wiki closeout/concept update
→ class-level skill patch if reusable procedure changed
→ memory only for stable facts
→ no runtime/cron promotion without separate approval

The decision at that time was not to write the article’s conclusion to memory, create an anthropic-dreaming skill, or start an automatic Dreaming cron job.

Follow-up option recorded by the 2026-05-11 / v0.13.0 snapshot

The historical assessment proposed that a later validation project could test a read-only weekly review job:

cron scheduled job
→ search recent sessions and project closeouts
→ generate candidate lessons only
→ deliver to an approved review channel
→ wait for explicit user approval before patching wiki/skills/memory

This would be an audited precursor to Auto Dream and should not mutate durable layers automatically in the first version.

Navigation

Type to search…

↑↓ navigate↵ selectEsc close