Patterns
Mubit was designed around a set of recurring agentic patterns — ways of using memory that make agents measurably better over time. Each page here names one pattern, explains when to reach for it (and when not to), and walks through it with real code against the SDK.
If you're new, start with the concepts section for the mental model, then come back here and pick the pattern that matches your problem. For complete end-to-end builds, see the Cookbook.
Learning from outcomes
The loop that makes agents improve: act, get judged, remember what the judgment means.
Credit the exact memories an agent used with the task's real result — the flagship Mubit pattern.
Distill a completed run into validated, reusable lessons.
Turn failures into preventive memory: rules, penalties, and distilled guardrails.
Version prompts and skills server-side and let outcomes drive their evolution.
Detect looping, drift, and stuck states — and break them without losing state.
Multi-agent memory
Coordination through durable, queryable memory instead of transcript passing.
Transfer work between agents as a structured handoff entry with a verdict loop.
Scopes, lanes, and gated promotion for a fleet that learns together — safely.
Orchestrator–worker with isolated contexts and a shared durable record.
Memory that survives and connects runs, users, and sessions.
Agents that wake on memory changes instead of polling.
Memory lifecycle
What a self-maintaining store does with knowledge over time.
New writes reconcile against existing knowledge — repetition strengthens instead of bloats.
Separate when something happened from when you learned it; supersede without deleting.
Idle-time merging and mental-model distillation, observable in your context blocks.
Bounded growth: unused evidence archives out reversibly; valuable memories stay hot.
Successful behavior becomes reusable workflows and provider-ready skills.
Context engineering
Getting the right memories into the window, at the right size, at the right time.