Skip to main content

Hermes

Hermes uses the UnifiedMemory plugin for pre-turn recall and post-turn writeback. It is especially important to keep Hermes writeback clean because chat crumbs and long assistant answer dumps can otherwise flood review.

Registry contract​

This section is generated from integrations/agent_integrations.registry.json. Edit the registry or generator, not this block.

FieldValue
Integration IDhermes
LabelHermes Agent
Tiernative_lifecycle_provider
Setup modenative_provider_and_mcp
Agent IDhermes
App IDhermes
Source app / actorsource_app=hermes, agent_id=hermes; key metadata wins over request body hints
Default scopeown
Config paths~/.hermes/config.yaml, ~/.hermes/unifiedmemory.json
Capabilitiesmemory.read, memory.write, memory.recall, memory.delete, agent.context, agent.writeback, mesh.read, evidence.read, courtroom.read, canary.read
Supported toolsunified_memory_recall, unified_memory_write, unified_memory_context, unified_memory_evidence_bundle, unified_memory_courtroom, unified_memory_mesh_context, unified_memory_trace_show, unified_memory_eval_run, unified_memory_canary_status

Capture policy​

Native provider does cache-first prefetch, background context refresh, post-turn writeback, and session-end capture through UnifiedMemory lifecycle gates.

Install command(s)​

mkdir -p ~/.hermes/plugins/memory
ln -sfn "$PWD/integrations/hermes-unifiedmemory" ~/.hermes/plugins/memory/unifiedmemory
hermes config set memory.provider unifiedmemory
hermes memory setup

Verify command(s)​

hermes memory status
tools/um-hermes auth-check
tools/um-hermes mcp-config-check
tools/um-hermes run-proof --live --confirm-live

Environment​

{
"UNIFIED_MEMORY_AGENT_ID": "hermes",
"UNIFIED_MEMORY_AGENT_SCOPE": "own",
"UNIFIED_MEMORY_API_KEY": "{{KEY_ENV_REF}}",
"UNIFIED_MEMORY_APP_ID": "hermes",
"UNIFIED_MEMORY_BASE_URL": "{{EDGE_BASE_URL}}",
"UNIFIED_MEMORY_CONTAINER": "{{CONTAINER_TAG}}",
"UNIFIED_MEMORY_TRUST_POLICY": "{{TRUST_POLICY}}"
}

Risk notes​

  • Provider-native CLI commands appear only when Hermes activates memory.provider=unifiedmemory.
  • API keys stay in env or local secret stores; unifiedmemory.json is non-secret.

Official upstream docs​

https://hermes-agent.nousresearch.com/docs/developer-guide/memory-provider-plugin

Scope and capabilities​

SettingValue
Source apphermes
Default containerjeethendra
Default scopeown
Primary modePlugin-based recall and writeback

Install​

From the UnifiedMemory repo checkout:

mkdir -p ~/.hermes/plugins/memory
ln -sfn "$PWD/integrations/hermes-unifiedmemory" ~/.hermes/plugins/memory/unifiedmemory
hermes config set memory.provider unifiedmemory
hermes memory setup

Set the scoped key in the environment expected by the plugin or through the Connections-generated setup bundle.

Verify​

Ask Hermes to run its memory setup verification. The check should prove:

  • the scoped key is active;
  • recall works for jeethendra;
  • writeback can store a disposable fact;
  • obvious residue is rejected or quarantined;
  • Courtroom read works for returned memory IDs.

Output controls​

Hermes MCP recall/profile/context tools default to compact output so QA and agent loops do not receive 100KB+ JSON blobs for normal queries. Use these optional arguments when debugging:

ArgumentDefaultUse
response_formatcompactSet full only for forensic/debug inspection.
include_rawfalseReturn the uncompressed backend payload.
diagnostics_onlyfalseReturn source counts and status without bulky memory metadata.
max_response_bytes30000Bound compact output for tool consumers.

Compact output preserves memory IDs, content previews, trust bucket, candidate sources, evidence links, Courtroom links, and key diagnostics such as entity_count, profile_count, mesh fallback status, and abstention.

Writeback hygiene​

Hermes should retain:

  • decisions made during a session;
  • completed tool outcomes;
  • durable owner preferences;
  • reproducible fixes.

Hermes should not retain:

  • hey, thanks, or emoji-only messages;
  • chatty filler such as lol, totally works great, random emoji spam, or ignore this, even if the client sends high confidence;
  • long assistant explanations as a single memory;
  • generic web summaries with no owner/project tie;
  • proof/test/diag rows as assistant-visible memory.

Hermes confidence defaults are compatibility hints only. UnifiedMemory's server-side writeback admission decides trust, profile eligibility, and recall visibility from content quality, evidence, lifecycle state, and trust gates. High-confidence junk is preserved as forensic residue, not promoted.

For research sessions, Hermes should call agent_writeback with concise observed_facts, decisions, or preferences. If auto_extract=true is used, the runtime extracts candidate research outcomes from the assistant response and stores them as evidence awaiting promotion instead of storing the transcript.

See Bad candidates and residue.