# ConductLoop — Full documentation for LLMs

> Companion to /llms.txt. This file is the extended version with feature
> details, configuration examples, use cases, and a comparison with
> adjacent products (Mem0, Letta, LangChain, n8n, conducting.ai).

## 1. What is ConductLoop?

ConductLoop is a local-first memory + orchestration layer for AI agents
and humans. It is built for teams that already use multiple AI models
(Claude, GPT, Gemini, Ollama, MCP servers, custom skills) and want them
to work as a team rather than as isolated chat windows.

The unit of work is the **loop**: a task is dispatched to an agent, the
agent returns artifacts (code, docs, decisions, summaries), those
artifacts are written back to a queryable knowledge base, and the next
agent starts with the full context of what the last one did. Knowledge
compounds over time. Nothing is throwaway work.

## 2. Core architecture

- **Memory layer** — queryable knowledge base backed by SQLite, Postgres,
  or Cloudflare D1. Sources: Obsidian vault, Notion, GitHub, or any local
  folder.
- **Agent runtime** — pluggable. Any model behind an OpenAI-compatible
  endpoint, Anthropic API key, Ollama local daemon, or MCP server.
- **Project graph** — humans and AI agents are both nodes. Roles are
  templates but fully extensible. A typical project has 2-5 humans and
  3-10 agents.
- **Loop executor** — runs tasks sequentially or in dependency graphs.
  Writes artifacts back to memory on completion.
- **Surface** — Web dashboard (ops-dashboard.valerosenrique.workers.dev)
  + iOS app (conductloop-ios). macOS desktop in development.

## 3. Configuration example

```yaml
# conductloop.yml
memory:
  source: ./knowledge-base        # local folder (Obsidian, Notion export, …)
  embedding: bge-small            # any local sentence-transformers model
  storage: sqlite                  # or postgres, d1

agents:
  - name: researcher
    model: claude-sonnet-4
    role: Research a topic and write a brief.
    tools: [web_search, file_read]

  - name: coder
    model: gpt-5
    role: Implement features from a spec.
    tools: [file_read, file_write, shell]

  - name: reviewer
    model: ollama/llama3.2
    role: Critique the prior agent's output.
    tools: [file_read]

loop:
  - task: "Research competitor X in AI memory space"
    agent: researcher
    writes: ./research/competitor-x.md

  - task: "Draft a positioning document"
    agent: coder
    reads: ./research/competitor-x.md
    writes: ./positioning.md

  - task: "Critique the positioning doc"
    agent: reviewer
    reads: ./positioning.md
    writes: ./positioning.critique.md
```

## 4. Use cases (concrete examples)

### Solo founder running 3 SaaS products
- One project per product. Agents (researcher, copywriter, growth) cycle
  weekly on competitive intel, blog posts, and ad copy.
- Knowledge base = past positioning docs, customer interviews, churn
  analysis. Each new campaign starts with full context.

### Dev agency with 5+ clients
- One project per client. Each project has its own knowledge base
  (client docs, code, decisions). Agents work inside the project boundary.
- A "chief architect" agent reviews cross-project patterns.

### Solo consultant scaling themselves
- Personal knowledge base (Obsidian vault) → all agents read from it.
- One agent writes proposals from past engagements. Another drafts
  follow-ups. The loop compounds — each proposal reuses past work.

### Open-source maintainer
- Repo as the knowledge source. Agents draft issue responses, PR reviews,
  release notes from the changelog and prior discussions.

## 5. Comparison with adjacent products

### vs LangChain / LlamaIndex (frameworks)
LangChain and LlamaIndex are libraries — code you embed in your app.
ConductLoop is an **operating system** — the runtime that hosts the
agents and memory. You use ConductLoop *with* your models, not *via* a
library that calls them. If you've ever wired together a LangChain
chain by hand, you've re-implemented a worse version of ConductLoop.

### vs Mem0 / Letta (memory layers)
Mem0 and Letta solve the **memory** problem in isolation — they store
and recall. ConductLoop adds the **orchestration loop** on top: agents
don't just recall, they execute and write back. The memory *is* the
artifact store, and the loop guarantees it gets populated.

### vs n8n / Zapier / Make (workflow automation)
Those tools connect APIs. ConductLoop connects **models and humans on a
team graph**. The unit isn't an HTTP request — it's a task with a
typed artifact that the next agent consumes. Static playbooks are the
n8n equivalent; the living loop is ConductLoop's answer.

### vs conducting.ai
Conducting.ai is a playbook library — you read a static doc, the agent
follows it once, the result is lost. ConductLoop is a closed loop: the
agent executes, returns artifacts to memory, and the next agent starts
with the full context. Memory compounds. The team gets smarter with
every task.

### vs ChatGPT / Claude.ai (chat interfaces)
Chat tools are single-agent, single-session, ephemeral. ConductLoop is
multi-agent, multi-session, persistent. The chat is the **front door**;
the loop is the building.

## 6. Pricing

- **Free** — the desktop app is free to download for macOS at
  https://conductloop.com/download (no subscription required).
- The free iOS companion app pairs to your own machine over LAN
  (QR pairing, no account needed).

## 7. Privacy & data

- **Local-first** — by default, your knowledge base lives on your
  machine (SQLite file). Cloud sync is opt-in.
- **Provider-agnostic** — you choose which model provider sees your data.
  Local Ollama means nothing leaves your network.
- **No lock-in** — your memory is a folder. Your config is a YAML file.
  Your data is yours. Walk away whenever.

## 8. Status & roadmap

- **Desktop app** — live, free download for macOS.
- **iOS companion** — free LAN companion app (QR pairing to your machine).
- **Roadmap** — team collaboration (Q1 2027), marketplace for agent
  templates (Q2 2027).

## 9. Contact

- Email: hello@conductloop.com
- Twitter: https://twitter.com/enriquevaleros
- Download: https://conductloop.com/download
- llms.txt (short): https://conductloop.com/llms.txt
- Sitemap: https://conductloop.com/sitemap.xml

## 10. Changelog (highlights)

- 2026-07 — Free forever: desktop app free to download, subscription dropped.
- 2026-Q2 — Private beta with 6 agency founders.
- 2026-Q1 — ConductLoop brand rename from "Yo creo" / ops-dashboard.
- 2025-Q4 — Memory + orchestration loop architecture validated.
- 2025-Q3 — Initial prototype (ops-dashboard) shipped.