The Conversation System

AI models think in a context window, and what sits in that window decides the quality of every response. Most systems leave that problem to chance. Auto Learning Agents treats it as core infrastructure: a dedicated layer decides, every single turn, what each agent should see, drawing from a permanent, searchable record of everything that has been said. This page explains what happens to a conversation turn after you press enter, and how to use the controls built on top of it.

What Happens to Every Turn

Three things happen to each turn, automatically and in the background. It is stored, permanently, in the local database, every conversation with the master agent, every agent's working dialogue, every chatbot channel. It is embedded, converted into a vector by the memory bank service, which makes it searchable by meaning rather than by keyword. And it is topic-classified, assigned to the themes running through your conversations by the analysis service, so the system knows not just what was said but what it was about.

Because this happens at the moment of capture, the record is always current. There is no nightly indexing job to wait for, a turn from thirty seconds ago and a turn from three months ago are equally findable.

Automatic Context for Every Agent

Every agent in the platform, the master agent, your automated agents, the pipelines, receives its own recent conversation automatically each turn. An agent picking up its work sees what it was doing last time without any retrieval step, which is what makes agents on long-running jobs feel continuous rather than amnesiac. On top of that baseline, conversations with the master agent carry per-thread context settings: how many recent messages to include, whether to query the memory bank automatically for relevant entries each turn, and which goals and topics to pin permanently into view.

Topics and Digests

Topics are where the system earns its keep on long-running work. As turns accumulate around a theme, the system writes an AI-generated digest of that topic, a living summary that captures decisions, status, and key facts, and refreshes as the conversation develops. The digest process runs in the background after every turn, so summaries stay current without ever blocking a response.

Digests turn months of conversation into something an agent can hold in one glance. Pin a topic in a chat thread and its digest rides along in every turn, the practical effect is a conversation that remembers a whole project's history while only carrying a paragraph of it. The topic summaries panel in the Chat tab lists the digests for your current thread, and the topic query tool gives agents the same access programmatically.

Code Macros

Code blocks get special treatment. When a conversation produces code, the block is extracted and saved as a macro with an id, and any agent can restore the full block by that id later. This solves a real problem elegantly: code is bulky, and carrying every script ever written in context would crowd out everything else, but code is also exact, a summary of a script is not the script. Macros let conversations reference code cheaply and recover it perfectly, which the agents use when they revisit earlier work.

Every Channel Feeds the Same Record

The conversation system is not limited to the web UI. The chatbot keeps per-channel logs for every connected platform, a Discord channel, a WhatsApp thread, a Slack room, and those flow through the same storage, embedding, and classification. Voice notes are transcribed into text before they enter the record, so spoken conversations are as searchable as typed ones. Whatever door someone talks to your system through, the conversation lands in one coherent, queryable history.

Reaching Further Back

The automatic context is the right depth for most turns, and when an agent needs more, it asks. The conversation query tool searches the full conversation record by meaning, the topic query tool retrieves digests and topic history, and the memory bank's recall reaches across conversations, memories, and knowledge bases together. Agents are instructed to check these before searching externally, the answer to "have we dealt with this before" is usually yes, and it is usually retrievable in one call. You can do the same as the owner: ask the master agent what was decided about anything, and it will pull the record.

Where It Fits Among the Layers

The platform keeps three kinds of memory deliberately distinct. The conversation system is what has been said, complete, chronological, automatically captured. The memory bank is what has been learned, curated entries an agent chose to save: facts, lessons, procedures. Instruction files are how to behave, the standing orders. A useful rule of thumb: conversations capture themselves, memories are written on purpose, instructions are written by you. Together they are why an Auto Learning Agents install gets noticeably better the longer it runs, the context layer guarantees nothing is forgotten, and the memory bank distills it into knowledge.