ClawLite v0.5.0

What's New in v0.5.0

Added

• /exportar (aliases: /export, /exporter, /exportieren) now delivers
  your complete memory as 7 Markdown files: chat history, profile (facts,
  goals, tasks, interests, reminders), conversation summary, brand
  profile, behavior patterns, knowledge graph, and watches. Previously it
  only delivered 2 files.
• Multilingual aliases for 5 commands that were hardcoded in Spanish:
  /exportar→export/exporter/exportieren, /olvida→forget/
  oublie/vergiss, /recordatorios→reminders/rappels/
  erinnerungen, /voz→voice/voix/stimme, /modelo→model/
  modele. Chinese/Russian/Japanese users are covered by the English
  alias (a real constraint of the Telegram Bot API: command names only
  accept Latin letters).
• The Synthesizer (used when combining research with brand/context) now
  distinguishes in its wording between findings corroborated by multiple
  sources and findings with conflicting reports across sources, reusing
  the same calibration mechanism already used by the research engine.

Fixed

• The "unknown command" message replied in English after a bot restart,
  regardless of the user's actual language — it read an in-memory session
  cache instead of the persisted language string already used by every
  other command.
• The /exportar zip's caption was hardcoded in Spanish — it now uses the
  multilingual catalog, same as every other deterministic message in the
  project.
• Removed a duplicate, concurrent write in hierarchical memory (behavior
  patterns and knowledge graph): two code paths were writing the same
  learning for the same turn, causing duplicate LLM calls and a real,
  unguarded race on behavior patterns.

ClawLite v0.4.0 — Automatic Ollama install + Docker step-by-step guide

ClawLite v0.4.0 – Initial Public ReleaseYour
personal AI assistant that finally respects the word “personal”.

ClawLite is a local-first, open-source AI assistant that runs entirely on your own machine. No data leaves your device unless you explicitly approve it. Built with real privacy and security by design — not as marketing features.

Key highlights:
One-click Windows installer — no terminal, no admin rights, no technical knowledge required
Runs locally via Ollama with transparent cloud fallback (always notifies you when it uses the cloud)
Strong security model: mandatory human approval for high-impact actions, Docker sandbox, and active defense against prompt injection
Persistent, private memory with semantic search that improves with use
Uses Telegram as the interface — no new apps to learn
Fully open source (Apache 2.0) — no subscriptions, no lock-in

This is the first public release. The focus has been on building a solid foundation with proper security and memory architecture.
See the full changelog for technical details.