← All projects
Security

SecOpentAgent

Self-improving AI agent infrastructure

Overview

SecOpentAgent is self-improving AI agent infrastructure — agent runtime with a closed-loop learning loop that creates skills from experience, improves them during use, and persists knowledge across sessions. Deployed behind a single gateway process reaching Telegram, Discord, Slack, Signal, and CLI, with scheduled automations handled natively. Used to back SecSight from the SOC side and to power SecOpent case-DSL exploration.

Seven runtime backends including serverless persistence — runs on a $5 VPS, a GPU cluster, or serverless infrastructure that hibernates when idle and costs nearly nothing between sessions. Compatible with the open agentskills.io skill standard and the Honcho dialectic user-modeling dialect.

Product impact

Pain → Solution → Value

Most AI chat sessions reset to a blank slate — skills lost between conversations, vendor lock-in everywhere, no way to drive the agent from your phone. SecOpentAgent turns that blank slate into a compounding asset: a single agent that learns, runs anywhere, and keeps memory across every model and surface.

Pain
Skills lost between sessions
Next chat has no idea who you are.
Solution
Closed learning loop
Create + improve skills + memory.
↑" data-i18n-zh="">
Each session gets smarter
Pain
Vendor lock-in
Closed gardens trap your data.
Solution
LiteLLM routes any model
Anthropic / OpenAI / local in one config.
1 → N
Config, N vendors
Pain
Can't run from phone
Chained to a desktop.
Solution
Telegram / Discord / Slack gateway
Same agent, any messaging app.
7
Platforms, 1 agent
Pain
No persistent memory of who you are
Role, stack, preferences — all gone.
Solution
Honcho user model + FTS5 search
Cross-session search + user model.
↑100%" data-i18n-zh="100%">100%
Context retention

What changes for the team

A wider view of the same loop. Each card zooms into the corresponding column above.

Pain

Every chat starts from zero

  • Skills taught today are forgotten tomorrow — the next session has no idea who you are.
  • Vendor lock-in traps your conversations, context, and tooling inside one closed garden.
  • You're chained to a desktop — no way to drive the agent from your phone.
0
Context carried across sessions
Solution

A self-improving agent, anywhere

  • Closed learning loop: skill creation + improvement + memory + cross-session search + user modeling.
  • LiteLLM routes Anthropic, OpenAI, local models, and self-hosted weights behind one interface.
  • Reach the same agent from Telegram, Discord, Slack, WhatsApp, Signal, or CLI — even while asleep.
5
Learning mechanisms, one loop
Value

Smarter every session. Yours, forever.

  • The agent gets sharper with every session — accumulated skills and memory compound like interest.
  • No vendor lock-in: swap models, self-host, or take it offline — your data stays yours.
  • Control from anywhere: message the agent from your phone, laptop, or chat thread — same brain.
7
Surfaces, one agent

Product architecture

Every request threads through five participants.

Technical architecture

AI agent · NousResearch/hermes-agent fork. Click any node or relationship for details.

Capabilities

Closed-loop learning
Agent-curated memory and skill self-improvement across sessions.
Multi-platform gateway
Single gateway to Telegram, Discord, Slack, WhatsApp, Signal, and CLI.
Scheduled automations
Built-in cron — daily reports, nightly backups, weekly audits in natural language.
Serverless-friendly runtime
7 backends incl. serverless persistence — hibernates when idle.

By the numbers

  • Closed-loop learning
  • 7 runtime backends
  • Multi-platform gateway
  • Scheduled automations
  • Honcho user modeling
  • agentskills.io compatible

Try it, fork it, or hire me to extend it.

Source on GitHub · {p['license']} · questions to 286043314+echocc00@users.noreply.github.com