The Living Story Company
The Lighthouse
Landfall
Copperheart
Driftwood
My Robot Friend
In VRThe Rememberer
Explainable-by-Construction AI and the Between-Session Memory Problem in Psychotherapy
Lance Nabers, LPC · Living Story Clinical · The Living Story Company Manuscript draft v2 · 2026-10-23 · prepared for academic review (v1: 2026-07-21) Status: conceptual framework + proposed study designs. No outcome claims are made in this paper; three studies are proposed to earn them. v2 adds §2.5, the developmental path the instruments are placed on (the Awareness Continuum Framework), §2.6, the instruments as they are built today, a paragraph on natural measurement points in §4, and a seventh limitation. v1's argument, its one commitment, its three principles and its six-step workflow are unchanged.
Abstract
Psychotherapy has a memory problem. A clinician typically observes one hour of a client's week — under 1% of lived experience — and reconstructs the remaining 167 hours from retrospective self-report at the following session, a process subject to well-documented recall and salience biases. Emerging artificial-intelligence tools promise to widen this aperture, but the dominant paradigm — opaque large language models producing post-hoc "explanations" of their own inferences — imports a second problem: explanations that research increasingly shows are plausible narratives rather than faithful accounts of model computation. This paper proposes a different architecture, XAI Therapy (Explainable AI–Assisted Therapy), organized around a single design commitment: do not explain a black box; never build one. In this framework, between-session data is generated by clinical instruments whose therapeutic logic is embedded in their structure (e.g., a cognitive-restructuring tool that natively captures situation, automatic thought, distortion class, evidence, reframe, and pre/post belief ratings), delivered inside narrative environments designed to sustain engagement, organized by AI systems restricted to citation-bound summarization, and released to care only through a licensed clinician's review. Explanation is thereby a property of the data's native structure rather than an after-the-fact account of a model's reasoning. We describe the framework's three principles (transparency, explainability, interpretability), its six-step clinical workflow, and its human-in-the-loop constraint; we place its instruments on a client-facing developmental path, the Awareness Continuum Framework (nine phases in four groups, walked as five weeks, with the clinician's service described in seven phases); we work a concrete example; and we propose three studies — a single-case experimental series, a documentation-burden trial, and a pilot randomized comparison — through which the framework's claims should be tested before they are asserted. The contribution is architectural and ethical: a falsifiable standard for how AI may enter the therapy room without asking anyone — clinician, client, or board — to trust what cannot be shown.
Keywords: explainable AI, psychotherapy, measurement-based care, ecological momentary assessment, human-in-the-loop, digital therapeutics, informed consent
1 · Introduction: The Fifty-First Date
There is a popular film in which a man falls in love with a woman whose memory resets every night; each morning he must win her again from nothing. Clinicians laugh at that film differently than other people do, because the weekly session is quietly the same scene. The client arrives; the week that has passed exists only as what can be recalled, in the order distress makes it recallable; the clinician asks how was your week? and receives — honestly, on both sides — a reconstruction. Then the hour ends, the door closes, and the richest data stream in mental health care, a human life as it is actually lived, goes unrecorded for another week.
This is not a caricature of poor practice. It is the structural condition of outpatient psychotherapy. The clinician directly observes well under 1% of the client's waking experience. The remainder arrives filtered through retrospective self-report, whose distortions — recency effects, mood-congruent recall, salience bias — are among the better-replicated findings in the field (Shiffman, Stone, & Hufford, 2008). Routine outcome measurement improves on unstructured recall and demonstrably improves outcomes, particularly for clients at risk of deterioration (Lambert & Shimokawa, 2011), yet remains under-adopted, in part because it adds administrative burden to clinicians already spending substantial time on documentation.
The author has a private name for this condition, borrowed from another film about a man reliving one day: the Groundhog Day problem. Therapy relives its own day. Each session begins by re-establishing what should never have been lost.
Artificial intelligence is now being offered as the answer, and the market's dominant answer is a listening machine: record the session, let a large language model infer what mattered. Whatever its utilities, this paradigm has a defect that this paper regards as disqualifying in a clinical context: the model cannot faithfully explain itself. When a clinician asks such a system why it flagged rising anxiety, the answer produced is a fluent rationale generated by the same opaque process that produced the flag — and a growing empirical literature finds that such self-explanations are frequently unfaithful to the model's actual computation (Madsen, Chandar, & Reddy, 2024; see also the broader survey literature on LLM explainability). The clinician is thus handed an insight she cannot verify, wrapped in an explanation she cannot trust, concerning a client to whom she owes informed consent. Rudin's (2019) argument for high-stakes domains applies with full force: where decisions matter, stop explaining black boxes and use models that are interpretable in the first place.
This paper takes Rudin's imperative as a design brief for psychotherapy and adds a clinical mechanism the interpretability literature lacks: the instrument. Psychotherapy already possesses, in its worksheet and protocol traditions, a century of devices that render inner experience into structured, theory-laden, clinically legible data — the thought record, the exposure hierarchy, the behavioral activation log, the self-efficacy rating (Beck, 2011; Bandura, 1977). These instruments are, in the machine-learning sense of the word, interpretable by construction: their output means something because their structure embodies the theory that gives it meaning. What they have historically lacked is a reason for clients to complete them and a channel by which their data reaches the clinician in time to matter.
XAI Therapy is the proposal that these two failures — the black box's opacity and the worksheet's abandonment — solve each other.
2 · The Framework
2.1 One commitment
We do not explain a black box after the fact. We do not build one.
Every downstream property of the framework follows from this refusal. If between-session data is generated by instruments whose clinical logic is structural, then explanation is not a report about a model's hidden state; it is the native shape of the record. The system's AI components are then restricted to roles that do not reintroduce opacity at the point of decision: organizing, computing over, and summarizing structured data with citation to it, and drafting clinician-facing syntheses that a human reviews, amends, or discards.
2.2 Three principles
Transparency — you can see how it works. The system's mechanisms, data flows, and the role AI played in building it are disclosed to clinician and client alike. (This manuscript itself discloses: portions of the software ecosystem described here were co-authored with AI assistance, openly.)
Explainability — every insight has evidence. An insight surfaced to a clinician must trace to (a) identifiable client-generated data points and (b) a named clinical framework. The system is prohibited from presenting an inference whose provenance it cannot display. Critically, this is a claim about evidence chains, not about model internals — the framework deliberately does not assert that any neural network's reasoning has been made faithful, a claim the current literature will not support.
Interpretability — the data means something clinically. A belief rating that falls from 70 to 30 across an evidence-examination exercise is legible to any CBT-trained clinician in seconds, because the number lives inside the theory that defines it. Interpretability is achieved at the instrument layer, where it is cheap and honest — not at the model layer, where it is costly and, at present, unreliable.
2.3 The six-step workflow
- The story invites. Instruments are embedded in narrative environments — in the present implementation, three interlinked story-worlds — because an unengaging instrument produces no data at all. Engagement is treated as the delivery system for measurement, drawing on the bibliotherapy and therapeutic- storytelling traditions (cf. Scogin et al.'s work on guided self-help) and on the observation that narrative assignment ("read chapter seven before next session") has always been a legitimate clinical act.
- The instrument does clinical work. Each tool operationalizes an established framework (cognitive restructuring, graded exposure, behavioral activation, self-efficacy building, grounding, paced respiration).
- The data is glass. The record produced is structured, framework-tagged, timestamped, and legible without inference.
- AI organizes, in the open. Software aggregates, computes trajectories, flags patterns (including the deferral signal — §3), and drafts summaries, each element carrying citations to the underlying records.
- The clinician decides. No AI-generated synthesis reaches the client or the record without review by the licensed human, who can inspect every citation. Human-in-the-loop is the architecture, not an override switch.
- The client keeps the story. The record belongs to the client — visible, portable, and exportable — on the ethical ground that the data is their lived experience, and on the clinical ground that ownership of one's own narrative is itself therapeutic.
2.4 What the framework refuses
The framework makes four standing refusals, stated here so they can be audited: it does not diagnose; it does not operate as a crisis service (safety resources are configured by the treating clinician for their own jurisdiction — a software default is somebody else's jurisdiction); it does not surface an insight without its evidence; and it does not permit any AI output to reach care without a human whose judgment becomes the record.
2.5 The path the instruments are placed on: the Awareness Continuum
The six-step workflow says how information travels. It does not say where a person is. For that, the framework places its instruments on a developmental path drawn from the author's clinical practice and named the Awareness Continuum Framework: nine phases in four groups, each phase carrying two names — the one a clinician uses, and the one a client reads.
| Group | Phases (clinical name · client name) | What rises |
|---|---|---|
| Foundation & Identity | 1 Identifying the Problem · Who are you and why are you here — 2 Road Mapping · Where are you going | Awareness |
| Mechanics & Deconstruction | 3 Habits, Skills · What are you doing? — 4 Self Assessment · The self assessment | Insight |
| Construction & Execution | 5 Building / Creating · Rebuilding, rewiring, reframing — 6 Driving the Car for the First Time · Your process, systems and loops | Intention |
| Actualization & Connection | 7 Integration Practice · Intentional purpose — 8 Meaning Making · Your definition of life — 9 Pass on to Others · Connected to the self and others | Purpose & Passion |
Underneath all nine runs the continuum the framework is named for: awareness increasing, then insight, then intention, then purpose and passion. In the author's words, the aim is to identify "what separated us from our self" and to get the person "back to that self."
Three rules keep the path from becoming a gate. First, a phase is where a person is, not a stage they must pass; every phase is open from the first day, and the deep work is assigned by the clinician whenever there is the strength for it. Second, a person's place on the path is set by the client and the clinician together, and where their two judgements differ, both are kept, with their dates. Third, the path is navigational: its order is a starting proposal, not a fixed sequence or treatment duration.
A person does not meet nine phases on a phone. In the story-worlds the path is walked as five weeks — Why now (phases 1–2), Notice life (3–4), Replace and reframe (5), Active practice (6) and Make it yours (7–9) — and under every week, every day follows one behavioural loop: learn something, do something, reward something, plan something (refine and repeat). The clinician's own work is described in seven phases (onboarding and informed transparency; therapeutic rapport; continuous assessment; behavioural model learning; task completion; skill building and generalisation; integration and ownership), each of which meets the nine where the work happens. The seven describe the service; the nine describe the person; the six steps describe how information moves between them.
2.6 The instruments as they are built
The present implementation has four parts, each built to the commitment of §2.1. The Lighthouse is the clinician's workspace: it is installed on the clinician's own computer and keeps the record there; its speech recognition and drafting models run on that machine by default, with a cloud model used only for the jobs the clinician allows, and every such job listed with where it runs. From a client's record the clinician can send any of up to 173 instruments; when the client uses one, a receipt returns to the record — started, finished, or still to come — limited to what the client has chosen to share. The note the clinician signs is the practice's record.
The instruments reach clients inside three story-worlds. Landfall, for one person, is thirty-five days on one island, walked as the five weeks above; its readings treat a change as real only when it is larger than the person's own day-to-day variability, a design choice intended to keep the software from reporting noise as progress. Copperheart, a recovery world, is thirty-two days told as a mystery at sea, each day's case solved with a recovery instrument. Driftwood, for couples and families, is thirty-five days on an island walked together, with exercises a couple or a family completes jointly. Each world can also be walked in a headset, against the same record. A fourth, smaller instrument, My Robot Friend, is a companion a person raises through small daily rituals; its growth only accumulates, and absence never costs a person anything, as a deliberate counter-design to the attention-economy mechanics limitation (3) warns against.
3 · The Signal the Hour Cannot See: A Worked Example
Consider the instrument this framework treats as paradigmatic: the cognitive- behavioral thought record, implemented as an interactive tool. A client completes it midweek, in three minutes, on a phone:
| Field | Client entry |
|---|---|
| Situation | "Received a critical remark on my work in front of the team." |
| Automatic thought | "Everyone thinks I'm incompetent." — belief 70% |
| Distortion (selected) | Catastrophizing / mind-reading |
| Evidence against | "The same lead praised two other pieces this month." |
| Balanced thought | "One piece of feedback is information, not a verdict." |
| Belief, after | 30% |
No inference produced this record. Its explanation is its structure: any CBT-trained reader can reconstruct exactly what therapeutic work occurred, verify it against the client's own words, and act on it in session. When an AI layer later reports "cognitive restructuring practiced twice this week; mean belief reduction 40 points; both episodes followed workplace evaluation events," each clause carries a pointer to records like the one above. The clinician clicks through, or doesn't — but can, always, and that capability is the difference between an insight and a rumor.
The same architecture captures a class of signal that retrospective recall almost never surfaces: process metadata. Whether homework was completed the day it was assigned or deferred until an hour before session; whether a couple completed a joint exercise together or separately; the hour of night a grounding tool is repeatedly opened. In the present implementation, the deferral interval is recorded explicitly. A clinician who knows a skill was practiced Tuesday morning versus 11 p.m. the night before session is holding different clinical information — and it is information no one had to remember, because, in the words of one of the framework's own design documents, logs are honest or they are useless at sea.
The clinical hour, so equipped, changes shape. The session no longer begins with archaeology ("how was your week?") but with material ("I saw the thought record from Tuesday — tell me about the meeting"). The author's clinical experience — twenty-five years, including the design of some of the earliest co-responder programs pairing mental-health professionals with police patrols — suggests the anxious silence when a client "has nothing to talk about" is among the most wasteful moments in practice. The framework's hypothesis is that this moment disappears not because AI fills it, but because the week finally arrives in the room. This is, however, precisely the kind of claim this paper declines to assert and proposes instead to test.
4 · Proposed Studies
The framework's promises reduce to three testable propositions. Each is paired here with a design deliberately scaled to a practice-based research setting, following the single-case experimental tradition (Kazdin, 2011) before the group-comparison one.
Study 1 — Signal: does explainable between-session data change the clinical picture? A single-case experimental series (N = 8–12 adult outpatients, multiple-baseline across participants). Baseline phase: treatment as usual with standard weekly self-report. Intervention phase: instrument-based between-session capture with clinician review. Primary measures: session-level rating of clinician-perceived information adequacy; count of clinically significant events identified in-session that were absent from client-initiated recall; client-rated session focus. Analysis: visual analysis plus Tau-U effect estimates across baselines. Falsification condition: if instrumented weeks do not yield reliably more clinically actionable material than recall weeks, the framework's core premise fails.
Study 2 — Burden: does structured capture reduce documentation load without degrading note quality? Within-clinician crossover (N = 10–15 clinicians, counterbalanced fortnights): notes composed with AI-drafted, citation-bound synthesis versus standard practice. Measures: documentation minutes per client (logged automatically), blinded note-quality ratings by independent reviewers on a standardized rubric, clinician-reported burnout inventory subscales. The marketing claim this study would earn — or forbid — is any statement of the form "reduces documentation time by X%." Until it runs, the framework makes no such claim.
Study 3 — Alliance and outcome: does the transparent architecture affect the things therapy is for? A pilot randomized comparison (N ≈ 60, two arms: treatment-as-usual vs. treatment plus XAI-instrumented between-session work) powered for feasibility rather than efficacy. Primary outcomes: engagement (instrument completion, deferral intervals), working alliance (WAI-SR), symptom trajectories on standard measures (PHQ-9, GAD-7) analyzed as slopes. Secondary: client-rated understanding of their own treatment — the informed-consent dividend the transparency principle predicts. Results would power a subsequent efficacy trial; no efficacy conclusion would be drawn from the pilot itself.
Natural measurement points. The worlds already hold a self-assessment rhythm the studies can use without adding burden: phase 4's self-assessment falls on day 13, every week closes by reading the week back (days 7, 14, 21, 28 and 35), and the evidence a person has gathered is read on days 13, 15, 21, 28 and 35. These points give Studies 1 and 3 within-person repeated measures at fixed positions on the path, and the clinician reads the same rows the client sees.
5 · Ethical Position
The framework's ethics are its architecture. Informed consent is only possible for a system that can be explained to the person consenting; non-maleficence is strained by any tool whose errors cannot be traced; clinical responsibility cannot be delegated to a process no one can audit. The human-in-the-loop constraint is therefore not a regulatory concession but the framework's central ethical claim: the licensed human's reviewed judgment is the only thing that ever becomes care. Data sovereignty follows the same logic in the client's direction — the record of a life belongs to the person living it.
A final disclosure belongs in this section rather than a footnote: this framework was developed by a practicing clinician in sustained collaboration with AI systems, and the collaboration is documented rather than obscured. The author takes the position that the field's credibility on AI transparency will be set by whether its builders practice it reflexively — beginning with their own papers.
6 · Limitations
Seven limits are acknowledged without defense. (1) No outcome evidence yet exists for the integrated framework; this paper proposes the studies that would create it. (2) Instrument-based capture samples what instruments capture; a life exceeds its worksheets, and the framework must not mistake the log for the person. (3) Narrative engagement mechanics could, poorly governed, shade into the attention-economy patterns this field should refuse; the framework's position is that engagement serves measurement and treatment, never itself. (4) The AI summarization layer, though citation-bound, still uses large language models whose drafting errors are possible; the human-review requirement is the mitigation, and its sufficiency is an empirical question Study 2 partially addresses. (5) Structured between-session visibility raises a surveillance risk if deployed coercively; the framework's client-ownership and clinician-review principles are designed against this, and implementation ethics must hold the line. (6) The author is the framework's originator and its first clinical user; independent replication is not a courtesy but a requirement. (7) The nine-phase path is a navigational design drawn from the author's clinical practice; neither its order nor its grouping has been validated, and the studies above should treat a person's phase as a covariate to be described, not an outcome to be claimed.
7 · Conclusion
The most human profession is being offered the least explainable technology, and the field's choice is not between adopting it and refusing it, but between two architectures: machines that must be trusted, and machines that can be shown. This paper has argued that psychotherapy already owns the solution in its oldest technology — the instrument, the structured exercise, the log a person keeps of their own change — and that the correct role of modern AI is to make that technology finally reach: reach the client midweek inside a story worth opening; reach the clinician before the session, organized and cited; reach the record as something a human chose to sign.
A century ago, the field learned that a thought written down and examined loses power over the person who wrote it. The claim of this framework is smaller than its ambitions and older than its software: what is recorded honestly can be worked with; what must be remembered alone is lost. The remembering, at last, can be shared — and every party in the room can see exactly how.
References
- Bandura, A. (1977). Self-efficacy: Toward a unifying theory of behavioral change. Psychological Review, 84(2), 191–215.
- Beck, J. S. (2011). Cognitive Behavior Therapy: Basics and Beyond (2nd ed.). Guilford Press.
- Kazdin, A. E. (2011). Single-Case Research Designs: Methods for Clinical and Applied Settings (2nd ed.). Oxford University Press.
- Lambert, M. J., & Shimokawa, K. (2011). Collecting client feedback. Psychotherapy, 48(1), 72–79.
- Madsen, A., Chandar, S., & Reddy, S. (2024). Are self-explanations from large language models faithful? arXiv:2401.07927.
- Rudin, C. (2019). Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead. Nature Machine Intelligence, 1, 206–215.
- Shiffman, S., Stone, A. A., & Hufford, M. R. (2008). Ecological momentary assessment. Annual Review of Clinical Psychology, 4, 1–32.
Additional citations to be verified and extended in v2 (bibliotherapy/guided self-help trials; measurement-based care implementation literature; LLM explainability surveys). Every reference above is real; none is decorative.
Correspondence: Lance Nabers, LPC · Living Story Clinical · ask@lancenabers.com This manuscript makes no claims of demonstrated efficacy. It is a framework paper with proposed designs, submitted in the conviction that the standard it argues for must apply first to itself.