AI Assistant memory
A practical explanation of what the AI Assistant remembers, where that knowledge comes from, and how it shapes every generation you run.
What memory is and why it exists
Every generation you run starts from scratch in terms of the AI Assistant conversation — but the AI Assistant does not start from scratch about your athletes or your coaching approach. Before generating anything, it loads two persistent memory profiles: one about you as a coach, one about the specific athlete. These are plain text summaries, built up over time from your previous work in Good Coach App.
Memory answers a simple problem: without it, the AI Assistant treats every athlete as a stranger and every coach as having no methodology. With it, the AI Assistant already knows that your athlete tends to go out too hard in tempo runs, or that you favour polarised training, without you having to state it each time.
Memory is not a database lookup. It is a synthesised text profile — a written summary of patterns and observations — that the AI Assistant reads as context before generating. Think of it as a briefing note that gets updated after each session.
The two memory types
Coach memory
Coach memory captures your coaching philosophy and methodology — the patterns in how you structure training that the AI Assistant learns from working with you over time.
What is stored:
- Your periodization philosophy: how you sequence phases, how you think about progression, what you consider non-negotiable in a training block
- Intensity distribution principles: what zone balance you favour, what you actively avoid and why
- Recovery and adaptation beliefs: your general stance on rest, how you think about loading and unloading
- How you describe and prioritise development goals for athletes
- What you consistently correct, request, or emphasise across multiple sessions
- Sport-specific preferences and methodology
- How you frame coaching decisions
What is not stored:
- Anything specific to one athlete's schedule, fitness level, or preferences — that belongs in athlete memory
- Specific numbers (weekly hours, session counts, volumes) that are determined by a particular athlete's situation rather than your universal approach
The distinction matters: coach memory should capture how you coach anyone, not the specifics of how you coach one person.
Athlete memory
Athlete memory is a persistent profile of an individual athlete — what the AI Assistant has learned about their physiology, behaviour, history, and response to training from your data and coaching notes.
What is stored:
Ongoing profile (synthesised over time):
- Performance trends and fitness progression, with rough timeframes
- Persistent health limitations, injuries, or recurring issues
- Training behaviour: consistency, tendency to over- or under-perform expectations
- Effort calibration: how well the athlete's perceived exertion matches objective data (HR, power, pace)
- Pacing behaviour: does the athlete typically go out too hard, negative split, or execute evenly?
- How the athlete responds to different workout types and intensities
- Goals, motivations, and any psychological tendencies relevant to training
- Context that is important for future planning but not recorded elsewhere
Recent workout log (rolling, last 3–5 entries):
- Date and workout type
- Whether subjective feel matched the data
- Execution quality: pacing, interval adherence, target achievement
- Unusual physiological signals (HR drift, power decoupling, poor recovery)
- Notable athlete comments if they reveal something meaningful
- Benchmark notes if the workout was a reference performance
When the log grows beyond five entries, the oldest are dropped and any lasting insights are folded into the ongoing profile.
What is not stored:
- General coaching principles or methodology — those belong in coach memory
- Anything that would apply equally to any other athlete
- Raw numbers without interpretive context — the workout files already contain those
Where memory comes from
Memory is not something you fill in manually. It is built automatically from the generations you run. Each source contributes a different type of knowledge.
Single workout analysis → athlete memory
When the AI Assistant analyses a single completed workout, it observes how the athlete executed it: effort calibration, pacing decisions, physiological signals, and how subjective notes compare to the data. After the analysis, athlete memory is updated with any new observations worth retaining.
This is the most granular source — session-level detail that accumulates into a picture of how the athlete typically performs.
Period analysis → athlete memory
When the AI Assistant reviews a broader training period (several weeks of data), it identifies trends that are not visible at the single-session level: consistency patterns, adaptation over time, how the athlete responds to loading across a block. After the analysis, athlete memory is updated with these medium-term observations.
Period analysis contributes strategic-level knowledge — the kind of insight that matters most when planning the next training block.
Weekly workout generation → coach memory
When you generate a week of workouts and include coach notes of substance (at least a sentence), the AI Assistant updates coach memory with any new signals about your methodology or preferences that emerged from those notes.
This is how coach memory learns incrementally from your natural coaching decisions — you do not need to describe your philosophy explicitly; it is inferred from what you consistently ask for and how you frame it.
Athlete memory is not updated by week generation. A new athlete with no logged analyses will have no athlete memory yet — the generator works from their profile data and your current notes instead.
Coach notes always override memory. If memory says "this athlete avoids tempo runs" but your note says "schedule a tempo run this Tuesday," the note wins.
Training cycles generation (after feedback) → both memories
After you accept or refine a cycles plan, both memories are updated:
- Coach memory receives any new signals about your periodization style, phase sequencing preferences, and how you structure long blocks — knowledge that only becomes visible when you plan at the macrocycle level.
- Athlete memory receives strategic-level context: the athlete's current fitness starting point, what the goals reveal about their training profile, any constraints or background you described, and what the generation conversation surfaced about where the athlete is and where they're headed.
This update happens only after the feedback pass — once you have confirmed or adjusted the plan — so memory reflects your actual coaching decisions, not just a first draft that may have been revised.
Where memory is used
Memory is read automatically before each generation. You do not need to activate it or reference it in your inputs.
| Feature | Coach memory used | Athlete memory used |
|---|---|---|
| Weekly workout generation | Yes — shapes session types, intensity balance, and how the week is structured | Yes — shapes volume, intensity, and progression decisions for this specific athlete |
| Single workout analysis | Yes — coach context informs how the analysis frames observations | Yes — prior behaviour and benchmarks inform interpretation |
| Period analysis | Yes — coaching approach informs how trends are evaluated | Yes — athlete history informs what counts as progress or a red flag |
| Training cycles generation | Yes — your periodization philosophy shapes phase structure and length | No — athlete-specific context should be stated explicitly in the goals description |
Training cycles generation is the one case where athlete memory is not injected. At the periodization level, the most relevant context is your coaching approach (already in coach memory) and explicit information about the athlete's current state and goals — which you provide directly in the goals description field. This is intentional: a cycles plan is built around stated goals and competitions, and the AI Assistant should not silently substitute remembered assumptions for your current brief.
What memory does not do
Memory does not replace your inputs. Memory reflects the past. If an athlete has had a recent injury, returned from a break, or significantly changed their fitness level, state it explicitly. The AI Assistant has no way to know what has changed since the last generation.
Memory does not store cycle structure. Training cycle dates, names, and phase sequences live in Good Coach App, not in memory. If a coach edits cycles manually after generation, memory is never updated to reflect that. Good Coach App is always the source of truth for what is currently planned.
Memory is not visible to you directly. There is currently no interface to read or edit the memory profiles. They exist as internal context. If you believe memory has accumulated something incorrect — for example, from an unusual period of training that is not representative — the best approach is to state the correction explicitly in your next generation's coach notes or goals description.
Memory does not enforce rules. Coach memory captures your typical approach; it does not prevent you from doing something different in a specific case. If you want to deviate from your usual methodology for one athlete or one training block, say so in your inputs — the AI Assistant will follow your explicit instruction over anything inferred from memory.
Memory in multi-coach clubs
When an athlete works with more than one coach — for example, a running coach and a strength and conditioning coach — the memory system handles each type differently.
Coach memory is always personal. Each coach has their own memory profile, built exclusively from their own inputs, notes, and generation decisions. The running coach's memory captures their endurance periodization philosophy; the strength coach's memory captures their approach to progressive overload and recovery. Neither coach's memory contains anything about how the other coaches. When each coach generates a plan, they bring their own methodology — not a club-wide one.
Athlete memory is shared. There is one athlete memory profile per athlete, regardless of how many coaches work with them. All coaches contribute to it and all coaches read from it. This is intentional: the facts stored in athlete memory — injury history, effort calibration tendencies, how the athlete responds to load — are true about the person, not about any specific coaching relationship.
The practical benefit in a multi-coach setup: knowledge flows between coaches automatically. If the strength coach's period analysis surfaces a recurring pattern (the athlete's performance drops significantly in week 3 of any loading block), the running coach will see that context when generating their next plan — without either coach having to communicate it manually.
What this means in practice:
- When you add an observation about an athlete in your notes or analysis, assume another coach may read it later. Context matters: "athlete struggles with sustained efforts above threshold" is more useful than an unqualified note, especially if one coach sees it in a running context and another in a cycling context.
- If another coach has recently updated the athlete's memory with observations from their discipline, you will benefit from that context automatically on your next generation.
- Athlete memory does not record which coach contributed which observation. It is a merged profile, not a log of who said what.
The shared nature of athlete memory is a feature, not a limitation — it mirrors how a good multi-disciplinary coaching team would brief each other before working with an athlete.
How memory improves over time
Memory compounds. Early generations will produce plans that reflect general periodization principles more than your specific approach. As you generate more analyses and cycles, the profiles become more accurate — and you will find the initial drafts require less correction.
An athlete with six months of workout analyses, several period reviews, and two or three completed cycles plans will have a rich profile. The AI Assistant will already know their injury history, how they respond to VO2max work, what their typical effort calibration looks like, and what the coach has been building toward — all without you restating any of it.
The practical effect: the more consistent your coaching activity in Good Coach App, the less you need to explain context. Memory turns your history into a standing brief.