Self-improving AI agent on C# / .NET 10
Creates skills from experience · improves them during use · remembers context between sessions

English · Русский

---

Hercules is a compact self-improving micro-agent that reproduces the key self-improving characteristics of "hermes-style" agents (Nous Research) in a runnable form factor.

The agent creates skills from experience, improves them during use, retains knowledge across sessions, and builds a deepening model of the user. It supports YandexGPT, Ollama Cloud, and Ollama Local through a single OpenAI-compatible interface (Microsoft.Extensions.AI).

This is the default landing page (English). For the Russian version, see README-RU.md.

---

✨ Features

Subsystem	What it does
Self-Improving Skill System	Automatically proposes creating a skill on repeated queries (>2 times), versions skills, and improves them when success rate is low
Long-term Memory	Stores user profile, preferences, entities, and session context in Markdown; carries context across restarts
Reflection Engine	Self-analysis after a session or every N commands — what went well / poorly / what to improve
Skill Router	Query routing: skill by triggers or direct LLM response
Hybrid Storage	Files (Markdown + JSON) for skills and memory + SQLite for logs and metrics
Multi-provider LLM	YandexGPT (primary), Ollama Cloud / Local, LM Studio — through a single OpenAI-compatible interface with automatic fallback
Interfaces	CLI (REPL, primary) + Telegram bot (secondary)

---

🏗️ Architecture

Hercules/
├── Program.cs                 # Entry point, DI and configuration setup
├── appsettings.json           # Provider and agent thresholds configuration
├── Config/
│   └── AppConfig.cs           # Configuration models
├── Agent/                     # Agent core
│   ├── AgentCore.cs           # Main request-processing loop
│   ├── SkillRouter.cs         # Skill-based routing (triggers)
│   ├── SkillManager.cs        # Skill CRUD + versioning (via LLM)
│   ├── ReflectionEngine.cs    # Self-analysis, reflection reports
│   └── MemoryManager.cs       # Long-term memory, user model
├── LLM/                       # LLM layer (Microsoft.Extensions.AI)
│   ├── ILLMClient.cs          # Unified provider interface
│   ├── ChatClientLLMClient.cs # Base over IChatClient
│   ├── YandexGPTClient.cs     # YandexGPT (OpenAI-compatible endpoint)
│   ├── LocalLLMClient.cs      # Ollama Cloud/Local, LM Studio
│   ├── LlmClientFactory.cs    # Client factory by provider name
│   └── ResilientLLMClient.cs  # Resilience + fallback chain
├── Storage/                   # Storage
│   ├── FileSkillRepository.cs # Skills/ — skill files
│   ├── MemoryStore.cs         # Memory/ — Markdown memory
│   ├── SqliteSessionStore.cs  # SQLite: sessions, logs, metrics, counters
│   └── Models.cs              # Domain models (Skill, SkillMeta, ...)
├── CLI/
│   └── ConsoleUI.cs           # REPL loop (Spectre.Console)
├── Telegram/
│   └── TelegramBot.cs         # Telegram bot (long polling)
└── Agent/WebApiAdapter.cs     # Core adapter for Web API + DTO

Additional "façade" projects on top of the core:

Hercules.WebApi/            # ASP.NET Core Minimal API (REST), port :5000
├── Program.cs                 # DI + CORS + middleware, reuses the core
├── Auth/ApiKeyMiddleware.cs   # X-Api-Key header validation
├── Config/WebApiConfig.cs     # API key + allowed CORS origins
└── Controllers/               # Chat / Skills / Memory / Stats (Minimal API)

hercules-web/                    # Astro + TailwindCSS frontend, port :4321
├── src/lib/api.ts             # Web API client
├── src/layouts/Layout.astro   # Base layout (dark theme, navigation)
├── src/components/            # ChatBox, SkillCard, ProfileEditor, StatsDashboard
└── src/pages/                 # index / skills / profile / stats

All runtime data goes into the data/ directory:

data/
├── Skills/                    # skill.{id}.md / .prompt.md / .meta.json / .usage.json / .v{N}.md
├── Memory/                    # user_profile.md, preferences.md, entities.md, context_{date}.md
└── sessions.db                # SQLite: sessions, interactions, metrics

---

🚀 Installation and Run

Requirements

• .NET 10 SDK

Build

cd Hercules
dotnet restore
dotnet build

Run CLI (primary mode)

dotnet run

Run Telegram bot

dotnet run -- --telegram

(set Telegram:BotToken in appsettings.json beforehand)

Run Web API (REST server)

# from repository root
dotnet run --project Hercules.WebApi

The server starts on http://localhost:5000. The agent core (AgentCore) is reused through the WebApiAdapter adapter — there is no separate agent logic in the Web API.

Run CLI via the main project

dotnet run --project Hercules -- --cli   # REPL mode
dotnet run --project Hercules            # same thing (CLI by default)

---

🌐 Web API

ASP.NET Core Minimal API. All responses are JSON (UTF-8, camelCase). Protection is the X-Api-Key header (value from WebApi:ApiKey, default dev-local-key). CORS is open for the local frontend (http://localhost:4321, http://localhost:3000). Every interaction is logged in SQLite (data/sessions.db).

Method	Route	Description
GET	/api/health	Liveness check (no key required)
POST	/api/chat	Send a message to the agent → response + mode/confidence/skill
GET	/api/skills	List skills
POST	/api/skills	Create a skill manually; ?ai=true — generate via LLM
GET	/api/skills/{id}	Skill details (metadata + prompt)
PUT	/api/skills/{id}	Update a skill (triggers/prompt/description) → new version
POST	/api/skills/{id}/improve	Improve a skill via LLM → new version
GET	/api/memory/profile	Long-term memory profile (Markdown)
PUT	/api/memory/profile	Overwrite the memory profile
POST	/api/memory/reset	Reset long-term memory
GET	/api/reflect	Run reflection → Markdown report
GET	/api/stats	Metrics: total, skill/direct, success rate, per-day

Example:

curl -X POST http://localhost:5000/api/chat \
  -H "X-Api-Key: dev-local-key" -H "Content-Type: application/json" \
  -d '{"message":"what is the weather in Moscow?"}'

Web API configuration (Hercules.WebApi/appsettings.json):

"WebApi": {
  "ApiKey": "dev-local-key",                  // empty string → no-key access
  "AllowedCorsOrigins": [ "http://localhost:4321", "http://localhost:3000" ]
}

---

🎨 Web Interface (Astro)

Minimalist SPA on Astro + TailwindCSS (dark theme, monospaced code blocks). Located in the hercules-web/ directory.

Page	Purpose
/	Chat with the agent (mode/confidence/provider badges, typing effect, skill hints)
/skills	Skill list, manual creation and AI improvement, editing
/profile	Long-term memory profile editor + reset
/stats	Metrics dashboard, skill/direct ratio, daily activity, reflection

Components: ChatBox, SkillCard, ProfileEditor, StatsDashboard. API client — src/lib/api.ts.

Run the frontend

cd hercules-web
npm install
npm run dev        # dev server on http://localhost:4321

The backend address and key are configured via environment variables (hercules-web/.env file):

PUBLIC_API_BASE=http://localhost:5000
PUBLIC_API_KEY=dev-local-key

Full local run (two terminals)

# Terminal 1 — backend
dotnet run --project Hercules.WebApi      # → :5000

# Terminal 2 — frontend
cd hercules-web && npm run dev                  # → :4321

Open http://localhost:4321.

---

⚙️ Configuration (appsettings.json)

{
  "Llm": {
    "Provider": "yandexgpt",                 // active provider
    "Fallback": ["ollama-cloud", "ollama-local"], // fallback order
    "YandexGpt": {
      "Endpoint": "https://llm.api.cloud.yandex.net/v1",
      "ApiKey": "<IAM or API key>",
      "FolderId": "<Yandex Cloud folder id>",
      "Model": "yandexgpt",                  // becomes gpt://{folderId}/yandexgpt/latest
      "Temperature": 0.6,
      "MaxTokens": 2000
    },
    "OllamaCloud": {
      "Endpoint": "https://ollama.com/v1",
      "ApiKey": "<Ollama Cloud key>",
      "Model": "gpt-oss:120b"
    },
    "OllamaLocal": {
      "Endpoint": "http://localhost:11434/v1",
      "ApiKey": "",                          // no key required locally
      "Model": "llama3.1"
    }
  },
  "Agent": {
    "SkillCreationThreshold": 3,             // repetitions before proposing a skill
    "SkillImprovementThreshold": 0.6,        // success_rate threshold for improvement
    "SkillEvaluationWindow": 5,              // skill evaluation window
    "ReflectionEveryNCommands": 10           // auto-reflection every N commands
  },
  "Telegram": { "Enabled": false, "BotToken": "" }
}

Any setting can be overridden via environment variables with the HERCULES_ prefix, for example: HERCULES_Llm__Provider=ollama-local.

LLM Providers

All providers work through an OpenAI-compatible interface and the Microsoft.Extensions.AI abstraction (IChatClient). Supported providers:

• YandexGPT — primary (Russia). The model is passed as gpt://{folderId}/{model}/latest.
• Ollama Cloud — cloud fallback (https://ollama.com/v1).
• Ollama Local / LM Studio — local fallback (http://localhost:11434/v1).

If the primary provider is unavailable, ResilientLLMClient automatically switches to the next one in the Fallback list.

---

💻 CLI Commands

Command	Description
> text	Direct query to LLM with profile context
/skills	Show all skills (table)
/skills create "name"	Create a skill manually
/skills improve {id}	Improve a skill (new version)
/memory show	Show user profile
/memory reset	Reset memory
/reflect	Run reflection manually
/help	Help
/exit	Exit with context saving and final reflection

🤖 Telegram Commands

• /start — initialization
• /skills — list of skills
• /profile — what the agent knows about the user
• /reset — reset memory
• plain text — agent response

---

🔄 How the self-improving cycle works

1. Request → load profile and context from memory
2. Routing → find a matching skill by triggers (SkillRouter)
3. LLM response → with the active skill (skill-prompt) or directly (direct)
4. Logging → input/output/confidence/mode in SQLite
5. Skill threshold → if the request has been repeated SkillCreationThreshold times → propose creating a skill (with confirmation)
6. Improvement threshold → if success_rate < SkillImprovementThreshold → propose updating the skill
7. Memory saving → facts about the user, entities, preferences
8. Reflection → at the end of a session or every N commands

Principles

• Never stop learning — every session enriches memory or skills
• Explicit improvement loop — the agent itself proposes fixes
• Transparent — the user sees all creations/improvements
• Human-in-the-loop — skills are created only after confirmation
• Versioned — old skill versions are not deleted (skill.{id}.v{N}.md)

---

🧪 Acceptance Criteria Check

Criterion	How to verify
A skill is created automatically	Repeat the same query 3 times → the agent proposes a skill
A skill is used	After creation — the query goes through skill: ...
A skill improves	After a series of bad responses → a proposal to update
Profile is saved	Restart → /memory show remembers facts
Context is carried over	Session 1: fact → Session 2: agent remembers
Reflection runs	After /exit — Reflection Engine output

---

📦 Dependencies (NuGet)

• Microsoft.Extensions.AI + Microsoft.Extensions.AI.OpenAI — AI abstractions
• OpenAI — OpenAI-compatible SDK (YandexGPT, Ollama)
• Microsoft.Data.Sqlite — SQLite
• YamlDotNet — metadata parsing
• Spectre.Console — improved CLI
• Telegram.Bot — Telegram interface
• Microsoft.Extensions.Hosting / Configuration.Json — DI and configuration

---

📚 Documentation

All documents are available in two languages. The default links point to the English version.

Document	Description
docs/QUICKSTART-EN.md · RU	Quick start in a few minutes
docs/ARCHITECTURE-EN.md · RU	Core and interface architecture
docs/AGENT-MESH-EN.md · RU	Micro-agent mesh concept
docs/IOT-SCENARIOS-EN.md · RU	B2C/B2B IoT deployment scenarios
docs/ROADMAP-EN.md · RU	Planned features and milestones
docs/CONFIGURATION-EN.md · RU	Full settings reference
docs/API-EN.md · RU	REST Web API reference
CONTRIBUTING-EN.md · RU	How to contribute
CHANGELOG-EN.md · RU	Change history
SECURITY.md	Security policy
CODE_OF_CONDUCT.md	Code of conduct

---

🤝 Contributing

PRs and Issues are welcome! Before starting, please read CONTRIBUTING-EN.md and CODE_OF_CONDUCT.md. Report vulnerabilities via SECURITY.md.

---

📝 License

MIT © 2026 Victor Buzin.