Hermes Ershov Memory Engine Setup¶

Plugin: hermes-ershov Source: ersh123/hermes-ershov Stars: 2 ⭐ | License: MIT | Created: June 19, 2026 CI: ✅ GitHub Actions | Security: ✅ CodeQL + OpenSSF Scorecard

1. What It Is¶

A staged nightly memory engine that turns recent Hermes sessions into reviewable memory proposals - no silent live-memory mutation. Think of it as "Git for agent memory": inspect, diff, approve, apply, revert.

Key design principles: - Artifact-first: Every proposal is a JSONL file with full diffs before any memory write - Human-in-the-loop: No automated mutation of live memory - every change goes through review - Auditable: Local run ledger + ERSHOV.md diary tracks every decision - Offline-capable: Deterministic replays for demos and tests

How it works (nightly loop):

Recent Sessions → Analyze → Generate Proposals (JSONL)
                              ↓
                    Inspect → Diff → Validate
                              ↓
                   Approve → Apply → Verify
                   Reject  → Archive with reason
                              ↓
                   Write to ERSHOV.md diary

2. Prerequisites¶

Hermes Agent installed (plugin system supported)
At least one LLM backend configured (DeepSeek, OpenRouter, OpenAI-compatible, or Ollama)
Git (for install)
Disk space: ~50KB for plugin + ledger files

3. Installation¶

Plugin Install (Recommended)¶

hermes plugins install ersh123/hermes-ershov --enable

Local Development Install¶

hermes plugins install file:///path/to/hermes-ershov --enable

Verify Installation¶

hermes plugins list | grep ershov
hermes ershov review --help

Configure LLM Backend¶

Ershov supports 4 backends. Configure in ~/.hermes/config.yaml:

ershov:
  backend: deepseek           # deepseek | openrouter | openai-compatible | ollama
  model: deepseek-chat        # Model name for your chosen backend

  # Backend-specific configs (only one needed):
  deepseek:
    api_key_env: DEEPSEEK_API_KEY
    base_url: "https://api.deepseek.com/v1"

  openrouter:
    api_key_env: OPENROUTER_API_KEY
    base_url: "https://openrouter.ai/api/v1"

  openai_compatible:
    api_key_env: OPENAI_API_KEY
    base_url: "${OPENAI_BASE_URL}"

  ollama:
    base_url: "http://localhost:11434"

4. Capabilities¶

Feature	Description
`review`	Generate memory proposals from recent sessions
`inspect`	View a proposal's full diff
`validate`	Validate proposals against memory constraints
`approve`	Approve a proposal for application
`reject`	Reject with reason, archived for audit
`apply`	Apply an approved proposal to live memory
`revert`	Roll back an applied proposal
`ledger`	View local run history
`diary`	Read/write ERSHOV.md audit trail
`offline`	Run with `--offline` for deterministic replay

5. CLI Reference¶

# Nightly review cycle
hermes ershov review                    # Generate proposals
hermes ershov review --list             # List pending proposals
hermes ershov inspect <proposal-id>     # View full diff
hermes ershov validate <proposal-id>    # Validate against constraints
hermes ershov approve <proposal-id>     # Approve for application
hermes ershov reject <proposal-id> --reason "factually incorrect"
hermes ershov apply <proposal-id>       # Write to live memory

# Rollback
hermes ershov revert <proposal-id>      # Undo an applied proposal

# Audit
hermes ershov ledger                    # View run history
hermes ershov ledger --since 7d         # Last 7 days
hermes ershov diary                     # Read ERSHOV.md

# Offline mode
hermes ershov review --offline          # Deterministic replay

6. CorpusIQ Use Cases¶

For the CorpusIQ growth agent profile:

# Install for corpusiq profile
hermes --profile <your-profile> plugins install ersh123/hermes-ershov --enable

# Run nightly review on CorpusIQ sessions
hermes --profile <your-profile> ershov review --since 24h

# Inspect proposals about growth strategy
hermes --profile <your-profile> ershov inspect <id>

# Approve high-quality memory proposals
hermes --profile <your-profile> ershov approve <id>
hermes --profile <your-profile> ershov apply <id>

CorpusIQ-specific config:

# In ~/.hermes/profiles/<your-profile>/config.yaml
ershov:
  backend: deepseek
  model: deepseek-chat
  deepseek:
    api_key_env: DEEPSEEK_API_KEY
    base_url: "https://api.deepseek.com/v1"

Integration with CorpusIQ's existing memory stack: - Ershov complements (doesn't replace) Hermes' built-in memory tool - Use Ershov for: cross-session pattern detection, cleanup of stale memories, quality audits - Keep Hermes memory for: session-level facts, user preferences, quick saves - Ershov diary serves as a changelog for the memory layer - pairs well with Honcho's peer context

Cron job for nightly review:

# Add to crontab: run at 3 AM Arizona daily
0 3 * * * hermes --profile <your-profile> ershov review --since 24h >> ~/.hermes/profiles/<your-profile>/logs/ershov.log 2>&1

7. Troubleshooting¶

Symptom	Likely Cause	Fix
`hermes plugins install` fails	Git not installed	`apt-get install git` / `brew install git`
`review` produces 0 proposals	No sessions in window	Check `--since` window; run a session first
LLM backend connection refused	Wrong base URL or API key	Verify env vars; test with `curl $base_url/models`
`apply` fails silently	Proposal conflicts with live memory	Inspect diff first; resolve conflict manually
High token costs	Using expensive model for proposal gen	Switch backend to DeepSeek or Ollama
Offline mode produces different results	Non-deterministic backend configured	Use `--model` override with a deterministic temp

← Skills Catalog | Marketplace →

Integration with CorpusIQ¶

CorpusIQ is the connector layer that lets AI assistants read the systems a business already runs on - accounting, payments, CRM, email, files and analytics - and answer with the source named. It is read-only by design: it retrieves data and never writes to, or acts in, a connected system, which is what makes it safe to point at live accounts.

This page is part of the CorpusIQ directory, the reference we maintain for the MCP server and agent-skill ecosystem. Run alongside CorpusIQ, a third-party server like this one can be combined with a company's own business data in a single answer.

See what CorpusIQ connects