| name | hermes-observer |
| description | Layered observer running NousResearch/hermes-agent alongside native self-improving-agent. Provides cross-session FTS5 recall, bounded memory curation, background review daemon, and skill self-improvement. Does not replace native auto-memory -- runs independently and surfaces insights the native stack might miss. USE WHEN: reviewing cross-session memory, recalling past conversations, comparing memory systems, checking observer health, triggering background review, searching session history. Triggers on: "hermes", "session search", "cross-session recall", "memory review daemon", "what did we discuss", "observer status", "hermes observer".
|
| version | 0.1.0 |
| author | MARPA Design Studios / GRAFTKIT |
| allowed-tools | Bash, Read, Glob, Grep |
Hermes Observer
A second pair of eyes on memory and learning.
Hermes Agent (NousResearch/hermes-agent) runs as a layered observer alongside GRAFTKIT's
native self-improving-agent. It does not replace auto-memory. It runs its own loop independently,
catching patterns the native stack might miss.
Principle
"Does it break something? No? Keep it."
Commands
| Command | What it does |
|---|
/hermes:status | Observer health -- is Hermes running, last review time, memory usage |
/hermes:search <query> | Search past sessions via FTS5 + LLM summarization |
/hermes:review | Trigger background memory/skill review manually |
/hermes:diff | Compare Hermes curated memory vs native auto-memory |
/hermes:cost | Show observer token/cost tracking from the ledger |
Architecture
Hermes runs as a separate process, not embedded in the Claude Code agent loop.
Native Stack Hermes Observer
----------- ---------------
auto-memory <--> MEMORY.md (bounded: 800 tokens)
self-improving <--> background review daemon
trace2skill <--> skill self-improvement (patch)
(no equivalent) <--> session search (FTS5 + LLM summarize)
(no equivalent) <--> memory security scanning
What Hermes adds that native lacks:
- Cross-session recall -- FTS5 search over ALL past conversations
- Background review daemon -- autonomous memory curation every N turns
- Bounded memory -- forced quality through scarcity (800+500 token caps)
- Skill patching -- improves existing skills based on real usage
- Memory security -- scans for prompt injection in memory writes
What Hermes does NOT do:
- Modify
.claude/ files
- Replace auto-memory writes
- Interfere with the native agent loop
- Share memory stores
Installation
hermes-agent is GitHub-only (not on PyPI):
cd /tmp && git clone https://github.com/NousResearch/hermes-agent.git
cd hermes-agent && uv pip install -e .
Cost Tracking
All observer costs logged to docs/overgraft-ledger.tsv per the System Improvement Protocol:
- Input/output/cache tokens per session
- Background review overhead
- Session search (Gemini Flash) cost
- Monthly aggregate with budget cap
Integration with Native Stack
| Native Tool | Hermes Equivalent | Integration |
|---|
/si:review | Background review daemon | Hermes runs automatically, SI runs on demand |
/si:extract | Skill creation from review | Cross-pollinate via trace2skill |
/si:status | /hermes:status | Both report independently |
| auto-memory | MEMORY.md + USER.md | /hermes:diff compares outputs |
Research
Full architecture analysis and gap assessment:
/Users/ojeromyo/Desktop/ToolFactory/.claude/agents/scout-report/journal/2026-04-03-hermes-agent-research.md
Activation Criteria
Stays active as long as it:
- Does not break existing functionality
- Stays within tracked cost budget
- Produces at least one actionable insight per week
- Does not conflict with auto-memory writes
