Hermes Improvements Package

What Hermes Improvements Package does

Hermes Improvements Package is an enhancement suite for Hermes Agent that adds vector-backed memory retrieval, adaptive persona alignment, reasoning tracing and task classification. Version 3.0.0 hooks AIAgent.run_conversation, the entry point the README says is shared by the gateway, CLI, TUI, ACP and one-shot modes, so the per-turn logic runs on every surface.

To protect prompt caching, v3 stops changing the system prompt mid-conversation and instead prepends adaptive guidance, such as learned rules, observed user style, task complexity and pre-fetched memories, to the incoming user message, while the original input is stored through persist_user_message. Components include a VectorMemoryStore using sentence-transformers or a TF-IDF fallback, dynamic prefetch of one to three relevant memories, an AdaptiveSoul that learns corrections such as be shorter, an AdaptiveWorkflow that labels tasks trivial to complex, and a ReasoningTracer. The README says this avoids loading the whole memory file into every system prompt.

Key features

• VectorMemoryStore with configurable embedding model and TF-IDF fallback
• Dynamic prefetch of one to three relevant memories per prompt
• AdaptiveSoul that learns style preferences and corrections from the conversation
• AdaptiveWorkflow task classification from trivial to complex
• ReasoningTracer with uncertainty scores and source attribution
• Hook injection that the README says survives hermes-agent upgrades

When to use it

• Retrieve only relevant memories instead of sending the full memory file every turn
• Let the agent pick up corrections like keep it shorter without editing rules by hand
• Trace the steps and sources behind an agent's reasoning

Who it is for: Hermes Agent users with large memory files who want semantic memory retrieval and adaptive behavior.

How it fits with Hermes Agent

Built for Hermes Agent, attaching to its AIAgent.run_conversation turn pipeline across the gateway, CLI and TUI.

Note: It works by injecting a hook into Hermes's agent_init.py.

FAQ