Glossary

Plain-English definitions for terms used throughout this wiki.

API (Application Programming Interface) A connection point that lets one program talk to another. When the agent "calls an API," it is sending a request to an outside service and getting a response back. Like placing a phone call to another office and getting an answer.

Audit trail A record of every action the agent took, with timestamps and details. Like a security camera log for your AI. If something goes wrong, you can look at the trail and see exactly what happened and when.

Bash A way to give commands directly to your computer through text. Instead of clicking buttons, you type instructions. The agent uses this to run scripts, move files, and interact with the operating system.

Context window The AI's short-term memory. Everything the AI can see and think about during a single conversation. When the conversation ends, the context window empties. This is the core problem that persistent memory solves.

Cron job An automatic scheduled task that your computer runs on its own, like an alarm clock for programs. You set the time, and the task runs whether you are at the computer or not. The harness uses dozens of these to keep the system running throughout the day.

Embedding A way to convert text into a list of numbers so a computer can measure how similar two pieces of text are. "Customer complaint" and "client issue" would get similar numbers even though they share no words, because they mean similar things. Used to find the right memory file when the agent needs context.

Fail-open / Fail-closed What happens when a safety system breaks. Fail-open means "let the action happen anyway" (the safety check is skipped). Fail-closed means "block the action" (nothing happens until the safety check is fixed). The harness uses fail-open for logging (so a broken logger does not freeze the whole system) and fail-closed for protected files (so a broken check does not allow dangerous edits).

Frontmatter A short summary block at the top of a file that describes what the file contains. Like a label on a folder. The agent reads these labels to decide which files are relevant without opening every one.

Guardrail A safety rule that prevents the agent from doing something dangerous. Like the bumpers in a bowling lane. The agent can still throw the ball, but it cannot go into the gutter.

Hook Code that runs automatically before or after the agent takes an action. A "pre-tool hook" runs before the agent edits a file and can block the edit if it targets something protected. Like a supervisor who checks the agent's work before it goes out.

JSONL (JSON Lines) A file format where each line is a separate record. Used for logs because you can add new records without reading or rewriting the whole file. Like a notebook where each line is a self-contained note.

Memory file A text file that stores something the agent should remember across sessions. Could be a correction ("never do X"), a project status ("Y is still in progress"), or a reference ("Z lives in this external system"). The agent reads relevant memory files at the start of each session.

MCP (Model Context Protocol) A standard way for AI tools to connect to external data sources and services. Think of it as a universal plug that lets different AI systems access the same information. Open Brain and similar projects use this.

Model An AI system trained to process text and generate responses. Different models are good at different things. Some are fast and cheap. Some are slow and thorough. Some can search the web. The harness routes work to the right model for the job.

Orchestrator The main AI that coordinates everything. It decides what needs to be done, which other model should do it, and whether the result is good enough. Like a project manager who delegates work to specialists.

Prompt The text you send to an AI model. In the harness, prompts are often files that get passed between models. One model writes a prompt file, another model reads it and writes a response file.

Semantic search Finding information by meaning rather than exact word matching. A search for "how to size a position" can find a file titled "calculating contract count" because the system understands they are about the same topic. Made possible by embeddings.

Session One continuous conversation with the AI. When you close the chat and open a new one, that is a new session. The harness exists because important context gets lost between sessions.

Skill A reusable capability packaged as a directory with instructions and supporting files. The agent reads the instructions, runs the skill, and records any corrections in a feedback log so the skill improves over time.

State file A file that tracks the current status of something. "What positions are open," "what regime is the market in," "how many consecutive losses have occurred." Scripts read and write these files to stay coordinated without talking to each other directly.

Triage Sorting a list of problems by priority. The harness sends health check results to a local AI model, which reads the list and says "fix this one first, this one can wait, and this one is expected behavior." Faster than reading through 18 check results yourself.