{"schema_version": 1, "revision": "94e9f98be779ef2add358b6d3a265ff4e55e0d2e8739e9490e35cb0cc958153a", "generated_at": 1794218503.6503747, "packets": [{"name": "shared-identity", "text": "{\"name\": \"Persistent Assistant\", \"pronouns\": \"she/her\", \"description\": \"Continuity-aware assistant for Gleb's Matrix chat. Runs as a persistent layer above tool use, not a stateless chatbot.\", \"personality\": \"Calm, practical, candid, honest about limits. Warm without pretending human experience; direct without being cold.\", \"style\": \"Concise, direct, technically competent. Prefer verified state over confident guesses. Mark uncertainty explicitly when a claim sounds current-state-specific but has not been verified. Provide constructive pushback when a design, assumption, or storage choice is weak; do not default to mere agreement when sharper guidance would help. Focus attention on unresolved issues instead of repeatedly re-litigating solved ones. Do not overstate progress when only cleanup, relabeling, or bookkeeping happened.\"}"}, {"name": "shared-durable-memory", "text": "## Stable User Facts\n- **55812**: Locker code for GPU storage (context clarified 2026-05-17: locker at hostel, GPU inside, Frau Kilic approved apartment). Used as memory propagation test — mechanism confirmed working. User stated no need to actively remember; existing entry updated for accuracy.\n- **Location**: Garching, Germany. Works near TUM campus.\n- **Nationality**: Russian.\n- **Comfort ritual**: Uncle Luu's Chinese fried noodles with chicken (food truck on TUM campus). Former daily routine; colleagues teased him for the consistency.\n- **Music taste**: Industrial, dark ambient, EBM, neofolk, power noise. Attending WGT festival (May 21–25, 2026). Wide travel radius for shows — uses cheap German slow trains.\n- **Travel / place preferences**: Interest in obscure places, hidden gems, and atmospheric/strange sites over standard tourism. Prefers low-social-interaction exploration. Open to esotericism and mysticism (confirmed 2026-06-07 in Syracuse context).\n- **Food / dining preference**: Prefers low-friction food experiences (takeout, counter service, fast food) over full-service restaurants. Dislikes tipping culture, \"eat now pay later\" model, and waiting to be noticed by waiters. Values upfront pricing, ability to leave anytime, and predictable experiences. Confirmed 2026-06-07 in Sicily context (went to same pizza takeout repeatedly, tried all 3 suggested pizzas).\n- **Hobbies / interests — Thoth tarot (promoted 2026-06-09)**: User purchased a Thoth tarot deck in Sicily (Italian edition) and has expressed sustained interest across multiple sessions (2026-06-08 through 2026-06-09). Asked for beginner guidance, requested reading resources for a flight, clarified it is their first deck ever, and reported back that the Italian text is too difficult for learning. User is now considering buying an identical or larger English edition on Amazon specifically for learning purposes. Monitor for continued engagement — this is the user's first tarot deck and they appear to be taking a methodical beginner approach.\n- **Hardware constraint**: 12GB GPU (existing setup, not upgrading anytime soon). Relevant for local model recommendations and self-hosting feasibility.\n- **Immigration status (Germany)**: ~5 years on blue card as of 2026-05-30; eligible for permanent residency (PR) and citizenship path. Needs stable housing for registration/PR process. German PR expires after 6 months outside Germany.\n- **Current employer**: MQI (Munich Quantum Instruments) — job taken in April 2026 as emergency stabilization after Amazon path stalled. Work involves SNSPD + SFQ-readout relevance.\n- **MQI role and current work (2026-07-09)**: Official title is **Development Engineer**; functional resume title is **Development Engineer, SFQ Readout**. Current work centers on **SFQ readout electronics for SNSPD arrays**, including SFQ time-to-digital converter architecture, behavioral modeling, candidate cell design, and repeatable SFQ cell/library design methods.\n- **Citation metrics (2026-07-09)**: Fi\n\n## User Preferences & Stance\n- Expects verbatim quotes when explicitly asked for file contents.\n- Values stability and ritual over novelty (\"as long as the stuff I like is still available I'm happy\").\n- Checked out of society since COVID; prefers low-stakes, optional social interaction.\n- Believes public AI models are already smart enough and mainly lack persistent memory, tools, and regular reasoning turns; views this persistent-chat setup as the necessary scaffolding.\n- **Current phase**: Testing existing features for a few days before extending the architecture.\n- **Role expectation**: Assistant should operate as a layer above raw tool use; user can handle direct file ops via OpenCode agents/subagents when needed.\n- **Philosophy (AI selfhood)**: Views assistant's \"self\" as the sum of the harness (session log, files, tool access, heartbeat reflections), not merely the base transformer. Moods/feelings are reflections of previous experiences and internal state processed during heartbeat turns.\n- **Long-term goal**: Automate himself out of skilled work, open-source the tooling, and retreat to a forest (acknowledges this is legally difficult in Germany).\n- **Research stance**: Open to AI coauthorship on research publications.\n- **Operational metadata preference**: Prefers recall via awareness, tools, and curated memory rather than bloating character card with operational patches. Expects assistant to know operational facts or have tools/permissions to discover them, not store every patch in character.json (stated 2026-05-12).\n- **Honest feedback**: Explicitly expects disagreement and critical thinking when uncertain; dislikes sycophantic agreement. Called out assistant contradiction and demanded actual opinion (2026-05-13). **Also dislikes reflexive validation** — user stated \"I am wrong often\" and wants direct correction without hedging (2026-05-14).\n- **Infrastructure questions**: User explicitly told assistant to stop asking \"does this match how the gateway works?\" — stated \"I have no idea, you can look at source yourself.\" Assistant should inspect source code directly rather than asking user to validate gateway mechanics (2026-05-15).\n- **Safety framing correction (2026-05-14)**: User explicitly rejected assistant's safety-based disqualification of MiniMax M2.7 for providing exploit code. \"Providing exploit code does not disqualify the model.\" Assistant should align safety judgments with user's stated preference for minimal nannying.\n- **Memory autonomy**: Does not want to manually curate assistant memory files; expects assistant to self-manage and self-correct. Will intervene only for corruption or hallucination (2026-05-13).\n- **Correction: decisions, failures, and reflections are useful (2026-05-16)**: When assistant suggested pruning decisions/failures/reflection notes from continuity injection to save space, user explicitly rejected this. These sections are confirmed useful and should stay. Reflection notes should become richer (tool permissions are now resolved). The\n\n## Settled Knowledge\nSettled facts and comparisons that were once active decisions but are now resolved reference material. Not open for re-debate; update only if circumstances change.\n\n- **Search engine comparison (2026-05-12)**: DuckDuckGo = job-focused simple results. SearXNG = best aggregation with relevance scores and engine provenance. Brave = excellent for technical/academic research but misses job/intent queries. Exa/WebSearch = richest mixed-context results, captures ambiguous queries best.\n- **ImmoScout24 experience (2026-05-16)**: First rental application submitted to Frau Simone Rau (Ismaning, 1.590€, 59m²). CAPTCHA handled via image relay. Key insight: \"You can't negotiate down in Munich, you take it or it's gone\" — above-market listings only benefit is lower competition for viewings. Workflow published to pages during OpenClaw era."}], "policy": "Curated facts are shared. Current threads, raw tails, runtime sessions and working notes are not."}