15.3 Three-Layer Architecture: Gateway / Engine / Plugin

☤ "Three layers aren't a classification — they're three ways of looking at the same system."

Why Hermes' Architecture Matters

Hermes layers a Plugin subsystem on top of the traditional channel+loop+tool structure. Memory, tool-execution backend, self-evolving loop, and voice transcriber are all plugins — giving Hermes the same "swappable kernel" flexibility as DeepSeek Harness (Ch.17), just earlier and more Pythonic.

Layer 1: Gateway

Same shape as OpenClaw but Python + more channels (15+). Cross-channel persona blending merges one human's channels into a single session, and Honcho user modeling is injected explicitly (not implicitly).

Layer 2: Engine — 5 Subsystems

async def agent_loop(session, user_message):
    context = await context_manager.assemble(session, user_message)
    for step in range(MAX_STEPS):
        parsed = parse(await llm.stream(context))
        if parsed.kind == "final_answer": return parsed.text
        if parsed.kind == "tool_call":
            decision = await permissions.check(parsed.tool_call, session)
            if not decision.allowed: ...
            result = await tools.run(parsed.tool_call)
            context.append_tool_result(parsed.tool_call.id, result)
            # ⭐ Hermes-only hook: self-evolving evaluation
            await self_evolving.on_tool_use(session, parsed.tool_call, result, context)

The self_evolving.on_tool_use hook is Hermes' unique addition — every completed task may trigger skill creation/update.

Context Manager — pulls from plugin Memory Store

async def assemble(session, user_message):
    long_term = await memory_store.recall_relevant(user_message.text, top_k=10)
    working   = await memory_store.get_session(session.id)
    episodic  = await memory_store.get_recent_episodes(session.id, k=5)
    persona   = await honcho.user_model(user_message.from_)
    return build_prompt(...)

Layer 3: Plugin Subsystems

hermes runtime set modal    # switch execution backend
hermes runtime set local

Memory: FTS5 + BM25

CREATE VIRTUAL TABLE memory USING fts5(
  content, created_at UNINDEXED, source UNINDEXED, tags,
  tokenize = 'unicode61 remove_diacritics 2'
);
-- search:
SELECT content, bm25(memory) AS score FROM memory
WHERE memory MATCH ? ORDER BY score ASC LIMIT 10;

BM25 weights term frequency but down-weights common words (IDF) — precise keyword recall without a vector database.

Mapping to Chapter 8's Pillars

Hermes implements all six pillars plus two new ones: Self-Evolution and Honcho User Modeling.

Section Summary

Next section: 15.4 Core: Self-Evolving Skills Loop