The Machine Learning Layer

Large language models arrive smart but generic, and nothing about calling one twice makes the second call better. The platform's answer is a learning layer that wraps the models: it observes what works in your install, turns observations into training data and retrievable knowledge, and feeds both back into how agents operate. This page maps the pieces and how they compound.

The Always-On Services

Two Python services run permanently, supervised with backoff restart so they recover from anything. The memory bank service (port 9500) does embeddings and storage: every conversation turn, memory entry, and knowledge chunk passes through it on the way to being searchable by meaning. The analysis service (port 9510) does the live ML: FastText topic classification, which is how every turn gets its topics, and River anomaly detection, online learning that spots when something in the system's behavior deviates from its own normal. Both serve the whole platform, the conversation system, the memory bank, and the agents all ride on them.

The Learning Node

The learning node is the layer's caretaker, ticking every 10 minutes on deliberately mechanical work: monitoring training accuracy so degradation is caught early, curating training data so the models learn from clean examples, tracking anomalies over time, and keeping statistics on suggestion approval rates, how often what the system proposes is what you accept. Its discipline is economic as much as technical: the routine passes use no AI at all, and it spawns a single AI analysis call only when accumulated signals genuinely warrant judgment. Learning overhead stays near zero until there is something worth thinking about.

Outcomes and Rewards

Learning needs ground truth, and the platform collects it as outcome events: pieces of work logged with how they turned out, the reply that got a response, the post that earned engagement, the task that completed clean. On its rounds, Impulse extracts the patterns behind successful outcomes into rewards, distilled statements of what worked, reviewable in the Knowledge tab so you can see and shape what the system believes about its own performance. Rewards close the behavioral loop: agents consult them as accumulated tactical wisdom, the approaches with evidence behind them.

Custom agents join the loop the same way the presets do: the outcome logging tools are part of the shared tool layer, so an agent you wrote last week can log its outcome events, what it tried and how it went, and its track record becomes learnable data alongside everything else. Instruct your agents to log outcomes for the work you care about, and the learning layer treats them as first-class citizens.

Training on a Rhythm

The schedule node retrains the ML models every 12 hours against the curated data, keeping the topic classifier current with how your install actually talks and the anomaly models honest about what normal looks like now. The deepest cycle is optional and weekly: with trainOllama enabled, a local Ollama model is fine-tuned on your install's own memory data, producing, week over week, a model increasingly fluent in your domain, on your hardware, from your data, going nowhere.

Every cycle described on this page runs on your hardware against your local database. The training data is your data, the trained models are your models, and the learning loop completes without a byte leaving the install, the same self-hosting promise the rest of the platform makes, kept in the place where it matters most.

What Learning Looks Like in Practice

The compounding is visible from the owner's seat if you know where to look. Retrieval gets sharper: topics classify more precisely, so the right context and digests surface more often. The memory bank deepens: more learned procedures, more callable scripts, fewer first-time mistakes, the tenth task on any beat goes smoother than the first. The suggestions improve: approval-rate stats mean the system knows which of its proposals you accept, and the rewards steer agents toward the patterns that earn acceptance. And the anomaly layer quietly watches the whole show, raising what deviates. None of it requires administration, the layer maintains itself, and the Knowledge tab keeps its conclusions inspectable.