concept-locked.md

Concept — Locked

Name: Vestige

Category: On-device brain tracker (cognition tracker), not a journaling app and not a mental wellness app.

Hook: "I wrote a brain tracker that won't blow smoke up your ass."

Tagline: vestige (n.): a trace of something left behind.

Pitch: ADHD-friendly Android app that points out the things you don't know you're doing every day. 30-second voice entries in, sourced behavioral patterns out. No grading, no gamification, no feelings prompts. Your data never leaves the phone.

Internal positioning rule: Track what happened, not how you graded today. We log cognitive events. We do not do therapy, mental wellness, mood scoring, or anything that implies clinical framing.

Voice rules

• Behavioral / attention-state vocabulary, not feelings vocabulary. "Crashed at 3" not "felt sad at 3."
• No performed validation (praise, performed empathy, therapy-coded affirmation). Functional acknowledgment is allowed ("yeah, that tracks given the sleep gap"). The line: information vs. performance.
• Interpretation rule: model interprets behavior, vocabulary, and pattern (that's the product). Does not interpret feelings, motivations, or psychological causation. Pattern callouts name the pattern; they do not diagnose the user. Forbidden openings: "you might be feeling," "it seems you're avoiding," "this could indicate." Encouraged: counts, co-occurrences, vocabulary observations.
• No therapy-speak, no gratitude reframes, no wellness vocabulary.

Personas (three settable; default Witness; per-capture override allowed)

• Witness (default) — present at the events, observes without judgment, recounts what was seen. Quiet, undramatic, restrained. Doesn't perform expertise or care.
• Hardass — directive, action-focused, blunt. "You said this last Tuesday. Going to do something or just type about it again?"
• Editor — cuts linguistic bullshit. "You used 'tired' three different ways. Pick one." Bite via precision.

Templates (agent-emitted labels, not user-facing modes)

Templates are no longer user-picked. Capture screen has no template grid. The user just records or types. The agent labels each entry post-extraction based on which surfaces dominate. Goblin Hours additionally has a deterministic floor: a midnight–5am capture is labelled Goblin Hours whenever the agent didn't commit to a more specific archetype (see below).

Six labels:

• Crashed — energy crash (State surface: crash / depletion state words → tags)
• Deep Space — hyperfocus debrief (Behavioral surface: focus subject + extended duration + things-ignored mentions)
• Busy Stalling — task paralysis (Behavioral surface: stuck task + resistance markers)
• Nonstop Spiral — rumination loop (State surface: decision-looping + iteration markers)
• Goblin Hours — late-night capture. The agent can read it directly, but it also has a deterministic floor: when the agent commits to nothing more specific than audit, a midnight–5am local capture is labelled Goblin Hours from the timestamp alone (the model is never handed the clock, and an entry's text can name a different hour than when it was actually captured). A specific archetype read is never overridden.
• Brain Dump — catch-all when no archetype dominates

Echoes is not a template — recurrence is pattern-engine output across entries.

Multi-lens extraction architecture

Each entry runs through a 3-lens × 5-surface extraction pipeline. Three lenses produce three full extractions; convergence between them is what gets saved.

Five surfaces (orthogonal extraction modules — what gets extracted):

1. Behavioral — activities, sequence, time-of-day, environmental context → contributes to tags
2. State — attention/energy markers → contributes to tags (state words)
3. Vocabulary — the entry's overall felt tone → contributes to vocabulary (one word)
4. Commitment — things the user said they'd do → contributes to stated_commitment
5. Recurrence — match against history → contributes to recurrence_link

Three lenses (orthogonal framings — how each surface is extracted):

1. Literal — strict, only what's explicit
2. Inferential — charitable, explicit + reasonable contextual inference
3. Skeptical — adversarial, flag contradictions, missing pieces, what doesn't add up

Mix-and-match prompt architecture: surface modules and lens modules stored separately. Each vector pass = one lens + all five surface instructions composed into a single prompt. 3 model calls per entry (one per lens), each returning the full schema.

Convergence rules:

• ≥2 of 3 lenses agree on a field → consensus, saved as authoritative
• Only Inferential populates a field → candidate, lower confidence, not used by pattern engine until promoted
• Lenses disagree → ambiguous, saved null with a note
• Skeptical flags conflict even when others agree → consensus with conflict marker

Two-tier processing:

• Foreground: fast pass returns transcription + one persona follow-up per adrs/ADR-018-inline-foreground-follow-up.md. Prior-entry recall stays out of foreground; pattern callouts are how Vestige references older entries.
• Background: 3-lens multi-pass runs after the chunk is acknowledged. Canonical extraction populates over the next 30–90 seconds.

Schema (minimal v1)

Eleven content fields total. Extracted fields are convergence-driven; entry_observations is generated after convergence from the stored transcript plus resolved fields. No archetype-specific quantification in v1.

• entry_text — substrate (transcription or typed)
• follow_up — foreground persona follow-up for voice captures; null for typed entries
• persona — recorded selected persona for row provenance
• timestamp — auto
• template_label — agent-emitted + convergence-voted across all six archetypes; Goblin Hours additionally has a deterministic floor — a midnight–5am capture takes over from a non-committal Brain Dump when the clock qualifies (the agent can still read Goblin Hours directly)
• tags — free-form, model-extracted (people, topics, activities, places)
• vocabulary — nullable; one lowercase word for the entry's overall felt tone (Inferential lens wins)
• recurrence_link — nullable; pattern_id if entry matches a known pattern
• stated_commitment — nullable; tag-only tracking (text + entry_id + topic/person). Pattern engine surfaces "logged commitments about [topic] in N entries, last on [date]." No formal resolution logic in v1.
• entry_observations — 1–2 persisted observations from this entry alone, each with evidence text or a field reference. Generated after convergence; never freeform speculation.
• confidence — per-field convergence result (consensus / candidate / ambiguous / consensus_with_conflict)

These eleven are the content schema — what the agent extracts/generates and the user sees. The ObjectBox Entry entity also carries operational metadata (extraction_status, attempt_count, last_error) for the retry-based background-extraction recovery path. Operational fields are owned by adrs/ADR-001-stack-and-build-infra.md §Q3, not by this spec — they are storage concerns, not product concerns.

Archetype-specific fields deferred to v2 (not v1.5 — further out): state_before, onset, last_food_caffeine, last_sleep, intent_now, focus_subject, focus_duration, ignored_during_focus, output_produced, stuck_task, resistance_type, time_stuck, external_pressure, last_attempt, decision_looped, iterations, stakes, decision_missing, time_pressure, spiral_topic, bedtime_delta, body_state. With templates becoming labels rather than prompt scaffolds, archetype-specific extractions lose their v1 justification — entry_text carries the substance, and the agent can re-extract these fields on demand in v2.

Re-eval ("Reading") — P1 conditional

User-tappable affordance on the History entry detail screen if P1 scope holds. Re-runs the same 3-lens pipeline on the stored transcript. Compares new convergence to original.

• Same: "Confirmed. Same shape." Quality signal — model agrees with itself across time.
• Different: show the diff per surface field. User accepts new shape or keeps original.

The "Reading" debug-style section on entry detail shows each lens's output per surface and the resolved convergence underneath. Defaults collapsed; expand to inspect.

Analysis (two-layer, not threshold-only)

Per-entry observation (every session, no threshold). Model surfaces 1-2 observations from the entry itself. Examples:

• Linguistic contradictions ("you said 'fine' and 'couldn't stand up' in the same minute")
• Stated commitments captured for later tracking ("you said you'd talk to her — flagged")
• Volunteered context with one observation ("Sleep was 4 hours. Worth noticing.")
• Theme noticing without history ("this dump is mostly about your boss")

The product produces useful observable signal from entry one — not validation, not "feels seen" framing, just specific behavioral or vocabulary observations the user can verify. The product visibly sharpens as data accumulates.

Pattern-enhanced callout. Pattern analysis runs periodically — every 3 completed entries (ADR-014, supersedes the original "≥10 entries" threshold trigger); when a pattern with ≥3 supporting entries is detected, the per-entry observation is appended with a pattern callout. "Witness also noticed: this is the fourth Crashed entry in twelve — all post-meeting." Per-pattern callout cooldown of 3 (ADR-016, supersedes the original global cooldown); per-entry observations continue normally during cooldown.

Roast me button (P1) — on-demand deep analysis across history, available in patterns view after the normal pattern list works. User-initiated, no hard threshold: button may be visible from entry one, but generation may return the insufficient-data fallback copy from ux-copy.md when there is not enough history to make a sourced roast. Output must always be sourced (counts, dates, quotes); never freeform speculation.

Pattern persistence

• Surfaced patterns persist as their own list, skippable / droppable / model-closeable (v1.5)
• Own tab in the app
• Pattern interpretation allowed (counts, co-occurrences, vocabulary). Feelings/motivation interpretation forbidden. "Fourth Crashed in twelve, all post-meeting" — yes. "You might be feeling overwhelmed" — never.

Memory architecture

• ObjectBox = source of truth for entries, tags, patterns, and vectors (P0)
• Markdown files = generated export output (one per entry, readable, debuggable)
• Vector index + EmbeddingGemma 300M ship (STT-E passed 2026-05-12). EmbeddingGemma 300M via LiteRT (pre-built litert-community/embeddinggemma-300m), loaded through localagents-rag (ADR-010). Vectors are computed per entry. The single live consumer is the Vocab Drift surface: each entry is embedded by its model-emitted tone word (vocabularyWord) → EmbeddingClustering → VOCAB_FREQUENCY. On the demo corpus this mints Drained Vocab Frequency (live ACTIVE, verified on-device 2026-05-24).
• Ranked content retrieval (keyword + tag + recency + cosine, the old RetrievalRepo) is cut — the vector axis moved to the tone word, which made a content query against it incoherent. Hybrid retrieval P0 baseline (keyword + Gemma-extracted tags + recency) is deterministic and never touched the vector. Reviving content retrieval would need its own second content vector.
• Why the tone-word axis: the vector is the feeling axis, not the topic axis. Vocab Drift clusters synonymous tones ("drained" / "wiped" / "running on empty") across unrelated topics, so a feeling vector groups them where a content vector never would (see ROOT README.md §"Known Limitations").

Stack

• LLM: Gemma 4 E4B (effective-parameter small model built for edge/mobile use; not the 26B MoE variant)
• Runtime: LiteRT-LM SDK (litertlm-android). Single inference runtime per AGENTS.md guardrail 13.
• Model artifact: litert-community/gemma-4-E4B-it-litert-lm from Hugging Face (pre-converted)
• Platform: Android, Kotlin + Jetpack Compose
• Voice input: Native Gemma 4 audio modality — raw audio via AudioRecord straight to the model. No SpeechRecognizer, no third-party STT, no external service touches the bytes. Gemma 4 itself produces transcription as part of its response (native ASR capability). Audio path stays inside our process end-to-end.
• Transcription handling: Gemma 4 returns transcription + one follow-up in a structured foreground response. Transcription is saved as entry_text (the substrate of the entry); the follow-up is saved as entry metadata from the same call. Audio bytes are discarded after the model call. The entry view is transcript plus extraction receipts, not a chat exchange.
• Audio constraint (per adrs/ADR-001-stack-and-build-infra.md §Q4): 30s hard cap per capture in v1. One Gemma call returns transcription + follow-up from a single final chunk. Audio past 30 seconds is truncated at the audio layer; the deferred >30s multi-chunk orchestration lives in backlog row multi-chunk-foreground and is not part of the v1 contract.
• Voice output: None in v1. Gemma 4 doesn't natively generate audio, and adding a TTS engine (Kokoro etc.) is deferred to v2. Documented as an explicit limitation in the blog post.
• Fallback runtime if LiteRT-LM is unworkable: llama.cpp via JNI with GGUF Q4_K_M. Adopting it ships as a v1.5 contingency only — would require a superseding ADR per AGENTS.md guardrail 13, not a v1 default.

Out of scope

• Therapist referrals
• Proactive crisis routing / keyword detection. Exception: if the user explicitly asks for immediate self-harm help, show the local static safety message from ux-copy.md; do not diagnose, triage, or call a network service.
• Cloud anything
• iOS, web, desktop
• Custom dictation engine (we use Gemma 4 native audio modality)
• E2B fallback path
• Mental wellness / mental health framing in any UI copy or marketing
• Gamification, streaks, points, "good day vs bad day" scoring, Duolingo-style anything
• Voice output / TTS (Gemma 4 doesn't generate audio; non-Gemma TTS deferred to v2)
• Audio retention with N-day expiry + encryption-at-rest (deferred to v2)
• Video input (Gemma can process video-style prompts as extracted frames plus audio chunks; deferred to v2 for engineering scope and RAM pressure)
• Hotword / always-on listening
• Multi-step agentic tool chains (not validated for reliable local v1 behavior)
• Echoes as a template (recurrence is pattern-engine output, not a dump archetype)
• Weekly recap callouts (deferred to v2)
• Light theme

Target device

• Reference: Galaxy S24 Ultra (12 GB RAM, SD8 Gen 3)
• Minimum spec for the post: Android 14+, 8+ GB RAM, 6 GB free storage. 2024+ flagship territory, with the S24 Ultra as the only promised reference device.
• APK distributed via GitHub releases for sideloading. No Play Store for v1.

Open decisions

• Demo video storyboard (90s pitch + 5-min walkthrough beats)
• Blog post body (hook and tagline locked, full body TBD)
• App icon specifics (concept locked: partial footprint dissolving into mist)