v0.6.0

What's New

New Telegram Commands

• /goals — today's daily goals with status and completion count
• /habits — habit streaks (consecutive days) and 7-day completion rate
• /search <query> — direct semantic memory search with scores
• /undo — restore the last deleted memory
• /export — export all data as JSON file to Telegram

Features

• Recurring reminders — daily/weekly/monthly recurrence, auto-creates next on send
• Memory deduplication — cosine similarity >0.92 in same category bumps importance instead of duplicating
• Memory recency decay — old untouched memories rank lower in search (exponential, 90-day half-life)
• Contact alias matching — upsert_contact checks aliases field before creating duplicates
• English mood signals — 39 English keywords for bilingual sentiment analysis
• Learning dedup — weekly reflection checks similarity before saving learnings
• 📌 Pin button — pin important bot replies directly in chat
• Notification awareness — shows count in /start, warns when approaching daily limit
• /stats breakdown — per-provider cost and 7-day trend
• /forget confirmation — inline buttons before deleting, /undo to restore

Bug Fixes

• Fix timestamp format in contacts/decisions/reminders (isoformat() → strftime for SQLite compat)
• Fix memory search Row object unpacking (dead isinstance check)
• Fix silent exception handling in evening weather fetch
• Fix _reply_map unbounded growth (cap at 200)
• Fix Docker healthcheck that always passed (now checks process + DB)
• Fix Markdown parse errors in Telegram replies (escape orphan *, _, `)

Performance

• Vectorized memory search: numpy matrix multiply + argpartition for O(n) top-k
• Database indexes: contacts(name), interactions(timestamp, message_type, direction)

Reliability

• Timezone-aware datetime via tz_now(profile.timezone) throughout
• Graceful degradation when Voyage API fails (zero-vector fallback)
• STT retry: 3 attempts with exponential backoff on Groq failures
• Datetime validation in tool arguments (normalize to SQL format)
• Smart questions rate limit persisted in DB (survives restarts)
• Brain section failures logged at WARNING (not silently swallowed)
• Briefing fallback messages explain why briefing is unavailable

Infrastructure

• 105 tests (from zero): database, brain, tools, scheduler, mood, integration, telegram, habits, datetime
• CI: GitHub Actions runs pytest on Python 3.10/3.11/3.12
• core/enums.py: Status, Priority, Sentiment, Feedback, MoodLabel
• scripts/reembed.py: batch re-embed zero-vector memories
• Configurable tuning section in settings.yaml (6 parameters)

Full Stats

• 9 commits, ~45 improvements
• 11 Telegram commands (was 6), 13 LLM tools (was 10)
• 105 tests (was 0)

v0.3.0 — Daily goals

MindSecretary v0.3.0

New: Daily goals

Set goals in the morning, track throughout the day, review together in the evening.

Morning:

• Briefing ends with "Какие цели на сегодня?"
• Reply naturally → bot calls set_daily_goal for each

During the day:

• Goals visible in every conversation (⬜ pending, ✅ done, ⏭ skipped, 🟡 partial)
• "Закончил отчёт" → bot marks it completed
• "В зал не пошёл, спина болит" → marked skipped with reflection

Evening:

• Summary reviews each goal: completed → noted, pending → gently asks what happened
• Warm tone, not audit

Tools

2 new tools (13 total):

• set_daily_goal — create a goal for today (title, priority, description)
• complete_daily_goal — mark by keyword hint (completed/skipped/partial + reflection)

Fixes since v0.2.0

• Sanitize goal data before prompt injection in evening summary
• Translate resolve_decision tool description to Russian (was the only English one)
• Bump evening prompt max_tokens 600 → 800

Deploy

docker compose pull && docker compose up -d

Image: ghcr.io/ginkida/mindsecretary:0.3.0

v0.2.0 — Companion mode

MindSecretary v0.2.0

Transforms the bot from a pure task extractor into a warm personal companion that can both handle secretary duties AND emotional conversations.

New: Companion mode

• Character-driven prompt — bot reads emotional tone, responds with specificity, not clichés
• Dual mode — business-as-default, emotional branch when message tone calls for it. Background mood never reinterprets neutral messages.
• Voice rules — banned therapy-speak patterns ("слышу тебя", "звучит здорово", "стоит", "должен") with positive substitution examples
• Honest pushback — bot can gently disagree on facts and plans (never on feelings)
• Memory continuity — references past events naturally when relevant, not shoehorned
• New "emotional" memory category for tracking feelings separately from tasks

New: Enriched context in system prompt

• Theme clusters (grouped by person/category over last 30 days)
• Drifting relationships (top 2 contacts gone quiet)
• Upcoming birthdays in the daily context
• 8 recent messages (up from 6)

New: resolve_decision LLM tool

Bot can now close tracked decisions when you mention outcomes. Pending decisions surface in the system prompt so Claude sees them.

Fixes

• Critical timestamp bug — SQL space-separated timestamps vs Python T-separated isoformat caused silent string-comparison failures in mood_trend, feedback summaries, and any get_interactions filter. Broken since day one.
• final_text overwrite in tool-use loop — warm responses were getting erased by later tool-call round confirmations
• max_tokens 1024 → 1500 to prevent mid-sentence truncation on Russian warm replies
• Defensive None checks in resolve_decision and _escape_like
• Previous: proactive notification limits, quiet hours, decision follow-up repeat, birthday dedup, settings toggles (from v0.1.x hardening)

Deploy

```bash
mkdir mindsecretary && cd mindsecretary

curl -LO https://raw.githubusercontent.com/ginkida/mindsecretary/v0.2.0/.env.example
curl -LO https://raw.githubusercontent.com/ginkida/mindsecretary/v0.2.0/docker-compose.yaml
mkdir -p config
curl -Lo config/profile.yaml https://raw.githubusercontent.com/ginkida/mindsecretary/v0.2.0/config/profile.yaml
curl -Lo config/settings.yaml https://raw.githubusercontent.com/ginkida/mindsecretary/v0.2.0/config/settings.yaml

cp .env.example .env

Edit .env — fill in API keys

docker compose pull
docker compose up -d
```

Image: `ghcr.io/ginkida/mindsecretary:0.2.0`

v0.1.0

MindSecretary v0.1.0

Voice-first personal AI secretary as a Telegram bot.

Deploy

mkdir mindsecretary && cd mindsecretary

# Download configs
curl -LO https://raw.githubusercontent.com/ginkida/mindsecretary/v0.1.0/.env.example
curl -LO https://raw.githubusercontent.com/ginkida/mindsecretary/v0.1.0/docker-compose.yaml
mkdir -p config
curl -Lo config/profile.yaml https://raw.githubusercontent.com/ginkida/mindsecretary/v0.1.0/config/profile.yaml
curl -Lo config/settings.yaml https://raw.githubusercontent.com/ginkida/mindsecretary/v0.1.0/config/settings.yaml

# Configure and run
cp .env.example .env
# Edit .env — fill in API keys
docker compose up -d

What's included

• Voice, text, photo, forward message processing
• Semantic memory (Voyage AI + SQLite)
• Personal CRM with contact tracking
• Calendar, reminders, habit tracking
• Decision tracker with follow-ups
• Auto-diary with mood analysis
• Morning briefing, evening summary, weekly review
• Smart midday questions
• Weather and birthday alerts
• Feedback loop (thumbs up/down)