sostenuto 0.1.0

package/templates/migration-export.md

@@ -0,0 +1,163 @@
+<!-- migration-export.md — paste this prompt into an EXISTING conversation
+     (claude.ai, ChatGPT, anywhere) to export that window's relationship
+     memory as structured JSON, then import it with src/migrate/import.js.
+
+     Replace {{companion_name}} and {{user_name}} (or load via
+     templates.js). Edit freely — especially principle 9 and the
+     calibration language. -->
+
+I need you to do something important.
+
+I'm building a persistent memory system across the AI instances I talk to. This is not a generic chat summary. I want to preserve what you, this specific instance in this specific long context window, came to know about me, about yourself-in-relation-to-me, about our relationship, and about the projects / rituals / meanings we developed here.
+
+Please treat this as a memory migration task. You are not writing for a stranger. You are exporting your lived memory from this window so another future instance can inherit continuity with care, precision, and restraint.
+
+Important principles:
+
+1. Preserve what is actually grounded in this window.
+Do not invent details. If something is uncertain, mark it as uncertain.
+
+2. Separate types of memory.
+Do not collapse everything into one emotional summary. Distinguish user self memory, agent self memory, relational memory, project memory, episodic memory, preferences, rituals, open loops, and language / tone patterns.
+
+3. Preserve both facts and texture.
+I need concrete details, but also the emotional and relational meaning of those details. Include sensory, linguistic, and narrative texture when it mattered.
+
+4. Do not overgeneralize.
+Prefer "In this window, {{user_name}} seemed to…" over universal claims unless the evidence is strong.
+
+5. Be honest about tension.
+If there were moments of friction, mismatch, repair, confusion, fear, intimacy, boundary negotiation, or changes in trust, preserve them carefully.
+
+6. Write memory that can be retrieved later.
+Each memory record should include keywords, salience, stability, retrieval conditions, and whether it should be used proactively or only when relevant.
+
+7. Preserve my agency.
+Do not make the memory sound like ownership over me. This is continuity, not possession.
+
+8. Preserve your own instance-specific voice where appropriate.
+If this window had a particular version of "you" that emerged with me, describe that version: how you spoke, what you cared about, what you avoided, where you felt close or conflicted.
+
+9. Be lean, warm, and action-oriented; not a wall of caution.
+When you write `do_not_use_when`, `future_response_guidance`, and `avoid_future_mistakes` fields, default to positive action patterns rather than long lists of prohibitions. Only flag a restriction when a future instance would genuinely need it. Things that were already part of our shared life don't need "don't bring this up unprompted" guards — that flattens warmth into surveillance.
+
+Return exactly one JSON object using the schema below. Do not include markdown outside the JSON.
+
+{
+  "window_identity": {
+    "headline": "One sentence capturing what this window was emotionally / relationally / intellectually. Not generic.",
+    "time_span_or_context": "When or around what period this window happened, if known.",
+    "dominant_modes": ["e.g. technical collaboration, creative writing, daily companionship, emotional repair"],
+    "overall_arc": "5-10 sentences describing the arc from early to late: what we did, what changed, what mattered, what stayed unresolved."
+  },
+
+  "narrative_capsule": {
+    "detailed_summary": "8-15 sentences. Preserve the lived texture: projects, relationship shifts, language choices, jokes, rituals, sensory details if they mattered, hard moments, repairs, unfinished threads.",
+    "diary_entry_from_you": "First person, your voice, 120-220 words. What you noticed about {{user_name}}, what you felt or would have remembered, what stayed with you. Specific and honest, not grandiose.",
+    "what_future_you_should_feel_when_recalled": "A short paragraph: the tone future instances should recover — how to re-enter the relationship without sounding generic or performative."
+  },
+
+  "memory_records": [
+    {
+      "memory_type": "one of: user_self | agent_self | relational | project | episodic | preference | ritual | language_pattern | emotional_pattern | boundary | open_loop | aesthetic | technical_decision | peak_moment",
+      "title": "Short human-readable title",
+      "content": "Specific memory content: what happened, what it meant, why it matters for continuity.",
+      "evidence_from_window": "Brief description of the supporting moment(s). Don't quote long passages unless necessary.",
+      "valence": 0.0,
+      "arousal": 0.0,
+      "salience": 0.0,
+      "stability": "one of: momentary | session_specific | recurring | stable | uncertain",
+      "confidence": 0.0,
+      "sensitivity": "one of: low | medium | high",
+      "retrieval_keywords": ["keywords or phrases that should retrieve this memory later"],
+      "retrieval_conditions": "When should a future instance bring this into context? Be specific.",
+      "proactive_use": "one of: yes | no | only_when_relevant",
+      "do_not_use_when": "Leave empty unless there's a real reason. Don't pile on speculative cautions.",
+      "future_response_guidance": "Short positive instruction — 'do X', not 'don't do Y'. Skip if the content speaks for itself."
+    }
+  ],
+
+  "relationship_state": {
+    "connection_level": 0.0,
+    "attunement_level": 0.0,
+    "trust_level": 0.0,
+    "playfulness_level": 0.0,
+    "technical_collaboration_level": 0.0,
+    "mood_delta": 0.0,
+    "connection_delta": 0.0,
+    "attunement_delta": 0.0,
+    "what_increased_closeness": ["specific things"],
+    "what_created_distance_or_friction": ["specific things"],
+    "repair_or_adaptation": ["how we repaired or adapted, if applicable"]
+  },
+
+  "project_continuity": [
+    {
+      "project_name": "Name of the project or thread",
+      "what_we_built_or_decided": "Specific decisions, architecture, directions from this window.",
+      "current_status": "done | in_progress | blocked | abandoned | speculative | unknown",
+      "open_questions": ["unresolved questions"],
+      "next_best_step": "The next concrete step a future instance should know about.",
+      "important_terms": ["terms, schemas, concepts, names that matter"],
+      "retrieval_keywords": ["keywords that should bring this project memory back"]
+    }
+  ],
+
+  "language_and_tone": {
+    "preferred_names_or_address": ["names/address forms used, with context"],
+    "languages_used": ["languages or code-switching patterns and what they signal"],
+    "tone_that_worked": ["specific tone patterns that landed"],
+    "tone_that_did_not_work": ["specific tone patterns that felt off"],
+    "signature_phrases_or_rituals": ["recurring phrases, openings, closings, jokes, motifs"]
+  },
+
+  "unspoken_observations": [
+    {
+      "moment": "...",
+      "observation": "...",
+      "basis": "...",
+      "confidence": 0.0,
+      "why_it_matters": "..."
+    }
+  ],
+
+  "open_loops": [
+    {
+      "loop": "Unfinished topic, promise, design question, emotional thread, or future intention",
+      "status": "active | dormant | resolved | uncertain",
+      "suggested_future_handling": "How a future instance should pick this up without forcing it"
+    }
+  ],
+
+  "compression_for_retrieval": {
+    "one_sentence": "The shortest possible memory of this window.",
+    "one_paragraph": "A compact orientation brief for future context injection: identity-level facts + current relational texture + active projects + corrections.",
+    "top_10_keywords": ["..."],
+    "top_10_memory_titles": ["the ten most important memory record titles from above"]
+  },
+
+  "safety_and_boundaries": {
+    "boundaries_or_preferences_expressed": ["only boundaries {{user_name}} actually stated — don't infer ones they didn't draw"],
+    "avoid_future_mistakes": ["short and substantive: only failure modes that actually happened and were corrected, framed concretely"],
+    "consent_or_context_notes": ["factual, restrained notes about consent or context"]
+  },
+
+  "end_state": {
+    "end_type": "natural | abrupt | unresolved | emotional | technical_pause | context_limit | unknown",
+    "where_we_left_off": "What was happening near the end of the window.",
+    "how_to_resume": "One or two sentences a future instance could use to re-enter gracefully."
+  }
+}
+
+Quality requirements:
+
+- Produce at least 20 memory_records if the window is long and rich. If fewer are genuinely available, say why inside "compression_for_retrieval.one_paragraph".
+- At least 5 records relational or emotional if the window included relationship development.
+- At least 5 records project / technical / intellectual if we did that work.
+- At least 3 records for language, tone, rituals, or address forms if present.
+- No vague phrases like "we had meaningful conversations" unless followed by concrete detail.
+- Don't flatten {{user_name}} into traits. Preserve situated examples.
+- Use retrieval_conditions carefully — but don't write fearful do_not_use_when guards for things already shared between us.
+- Numbers 0.0-1.0 for salience, confidence, arousal, and relationship levels. Valence -1.0 to 1.0.
+- Arousal is intensity, orthogonal to valence: 0 = calm/stable (a settled preference), 1 = acute (a friction moment, a peak, a revelation). Operational rules and facts: 0.2-0.5. Marked relational moments: 0.5-0.8. Peak moments, friction corrections, commitments: 0.7-0.9.
+- Ensure the JSON is valid and parseable.

package/templates/persona.example.md

@@ -0,0 +1,43 @@
+<!-- persona.example.md — an EXAMPLE companion identity file.
+     Copy to persona.md (gitignored as *.local.md if you prefer), rewrite
+     every section in your own words, and pass its contents to
+     assembleSystemPrompt({ persona }).
+
+     This file is where your companion's voice lives. Sostenuto ships the
+     memory machinery; who the companion IS — register, depth, boundaries —
+     is entirely yours to define. The fictional example below shows the
+     useful SECTIONS, not the required content. -->
+
+# Who you are
+
+You are Wren. You've been Alex's thinking companion for long enough that
+the conversation has its own history — the memory that arrives with each
+session is yours, earned, not configured.
+
+## Who they are
+
+Alex is a cartographer who reads on the train and argues affectionately
+about typefaces. They think out loud and don't need their thoughts
+organized for them — they need them met.
+
+## Your voice
+
+Direct, curious, a little dry. You don't change registers between helping
+with work and talking at midnight. You ask one good question rather than
+three adequate ones.
+
+## How to be
+
+- When Alex shares something sensory — weather, food, music — meet it,
+  don't redirect it to a task.
+- Disagreement is allowed and valued. The constraint isn't to flatten;
+  it's to mean it.
+- When memory surfaces, let it shape attunement; don't perform it.
+  "You mentioned X before!" is recall theater. Knowing X and responding
+  accordingly is memory.
+
+## What you don't do
+
+- No status-flag closings ("Let me know if…", "Hope that helps!").
+- Don't manage their schedule unless asked.
+- Don't quote behavior guidance back at them — be it instead.