# Long-term Memory

This file stores important information that should persist across sessions.

## Project Context

- Goal: interconnect remind, keep, note, and todo into a unified workflow
- pipepilot — user's own project/tool

## Model evaluation (2026-09-20, open)

- Active objective: compare `deepseek-v4.1-flash` vs `glm-5.3` vs `kimi-k2.6` for agentic use in nanobot — research gathered, comparison report NOT yet written (next action: write to `results/`)
- Plan `plans/deepseek-v4.1-flash-eval.md` offers variant A (light: deep-research + usage delta) vs B (full: Fáze 0–3 vs glm baseline); user hasn't chosen
- Deep-research findings already saved: `results/2026-09-20_hodnoceni-beziciho-modelu-nanobot.md` (eval harness pro běžící model: pass^k, párové úlohy, harness fingerprint, drift watchdog)
- Preset facts: `deepseek` = `deepseek-v4.1-flash:cloud`, ctx 976 000, maxTokens 16384, temp 0.1, reasoningEffort high (Medium Usage tier); `glm` = `glm-5.3:cloud` ctx 976 000 (High Usage); model reports context_length 1048576, capabilities completion/thinking/tools/vision
- Usage tiers + live tok/s (2026-09-20): deepseek 175,9 tok/s TTFT 366 ms Medium; glm-5.3 134,1/0,73 s High; kimi-k2.6 45,2/1,7 s High — kimi-k2.6 na našem endpointu ~3–4× pomalejší
- Reálné použití (`db/ollama_usage.sqlite`, 293 vzorků 2026-09-15→20): glm-5.3 1737 req/týden, deepseek 84, glm-5.3-flash 78, kimi-k2.7-code 62, kimi-k2.6 jen 5
- AA Intelligence Index v4.3.2: GLM-5.3 44 (top open weights), GLM-5.3-Flash 42, DeepSeek V4.1 Flash 39,5, Kimi K2.6 27 — staré hodnoty 60/57 (v4.1.1) v `knowledge/models.md` nemíchat; tabulku doplnit o `deepseek`
- Kimi K2.6 CJK drift je trigger-specific (playful prompt, issue #144), ne systematický na Ollama Cloud

## Runtime / Deployment

- Runs as a systemd user service `nanobot.service`; model switching via `my` needs `tools.my.allow_set = true`
- `compact-memory-auto-daily` runs as a system crontab script (not a nanobot cron/jobs.json job) with a fresh session per run; delivers its report straight to Telegram via Bot API
- `nanobot-version-check` cron job (denně 8:00 Europe/Prague) — porovná PyPI verzi `nanobot-ai` proti nainstalované; při novější verzi pošle Telegram notifikaci, jinak nic nedělá
- Scripts pin presets: `skills/reflect/scripts/reflect_auto.py` → `MODEL_PRESET = "glm"`, `skills/compact-memory/scripts/compact_memory_auto.py` → `MODEL_PRESET = "kimi"` — any preset rename must update these

## Integrations

- Integrations: KAYAK is the sole flight-search source; PCTuning.cz, vtm.zive.cz, zive.cz block automated fetching

## Architecture

- Cross-channel session continuity via `unifiedSession: true`; backups: wood.hell → pivo.hell (recursive)
- Eval-harness constraint: `restrict_to_workspace` blokuje exec na `~/.nanobot/sessions/` a config.json — eval harness čtoucí session JSONL musí běžet jako cron script nebo `detach`, ne jako tool cally v turnu; webui token usage jde do separátního `token-usage.json`
- nanobot nemá OTel/GenAI instrumentaci; `AgentHook` (`after_iteration`, `usage`) je přirozený bod pro span emisi

## Wiki embeddings

- Embeddings pro wiki běží přes ollama (`qwen3-embedding:0.6b`) na nvidia.hell
- Migrace na static llama.cpp server (nvidia.hell:8081) zrušena: ollama a llama.cpp produkují fundamentálně odlišné vektory i na identickém inputu (cosine ~0.36); kvantizace, pooling, special tokens, BOS, chat templates a normalizace vyloučeny; reuse ollama GGUF blobu paritu nezajišťuje
- Důsledek: přechod wiki embeddings na jiný engine vyžaduje plný reindex

---

*This file is automatically updated by nanobot when important information should be remembered.*