Instruction Files

This is the third memory layer, alongside the conversation system (what has been said) and the memory bank (what has been learned). Instruction files are how to behave: the standing orders each agent carries into every tick. Keeping them as files has a compounding payoff, behavior is inspectable, editable, versionable, and changeable while the system runs, because agents read their instructions fresh each time they work.

The Layout

Everything lives under the memory directory of the install. memory/core_master.md is the master agent's role: how it addresses you, what it prioritizes, how it uses the system on your behalf. memory/core_impulse.md is the same for Impulse, the autonomous layer. memory/chatbot_channel.md shapes how the chatbot behaves in every connected channel.

memory/agent_systems/<type>/ holds a directory per agent type. Each one contains the task agent's own role instructions, the per-type guides the master agent and Impulse read when they deal with that area, and, for the pipeline types, the step files that script each pass of the coding and research pipelines. This is the deepest layer of customization: the presets are entirely defined here, and reading a preset's directory is the best way to understand exactly what it will and will not do.

memory/system/ holds the owner-editable settings files the UI exposes directly: your brand voice, product descriptions, support policies, coding standards, compliance rules, and image guidelines. memory/master/ is an optional shelf of reference documents you maintain for the master agent. memory/projects/ tracks ongoing work, living notes the agents keep as projects develop.

The System Files: Your Voice, Everywhere

The memory/system files deserve special attention because they cut across every agent at once. The brand voice file shapes the tone of every outgoing message, marketing email, support reply, social post. The product file is what agents consult to describe what you sell. The policy and compliance files set the boundaries customer-facing agents honor. The coding standards file steers every line the coding pipeline writes. Edit one paragraph in one file and the whole fleet shifts together, which is exactly how a small owner team keeps dozens of agents on-message.

These files are editable straight from the web UI, and filling them in is the single highest-value setup step after installation. An agent fleet with a rich brand voice file and honest product descriptions sounds like your company; the master agent's onboarding conversations help you write them.

Instructions for Your Own Agents

Custom agents point at their own instructions file through the instructions_file field in agents.json, and the file is the agent. A good instructions file reads like a brief for a competent contractor: what the job is, what done looks like, what tools to use, what to leave alone, and when to raise a flag instead of deciding alone. The creating agents guide walks through the structure, and the preset role files in memory/agent_systems are excellent worked examples to copy from, they are production instructions, not samples.

The master agent will happily draft instructions with you. Describe the job in chat, iterate on the draft together, and point a new agent at the result, the whole loop takes minutes, and the file remains yours to refine as you watch the agent work.

A useful skeleton for a first instructions file:

# Role
You monitor our product reviews and summarize sentiment.

# Each tick
1. Check for new reviews since the last run.
2. Save notable patterns to the memory bank.
3. Flag anything needing a human reply.

# Boundaries
Never reply to reviews directly. Flag instead.
Keep summaries under 200 words.

Short, concrete, and bounded beats long and aspirational. Agents follow what is written, so write the job, the rhythm, and the limits, then let the memory bank accumulate the judgment.

Why Files Win

Putting behavior in files rather than code is what makes the platform's central promise work: the AI models are replaceable workers, and the system is the structure around them. Because the role travels with the file, not the model, you can switch an agent from Claude to a local Ollama model and its behavior specification stays identical. Because files are read fresh, edits apply live. Because they are plain markdown, you can diff them, back them up, and review exactly what changed when an agent's behavior shifts. And because agents themselves can read these files, the system can explain its own behavior, ask the master agent why a preset acted some way, and it can quote the instruction that drove the decision.