sillytavern-character-memory/CHANGELOG.md

Changelog

1.2.1

Bug Fixes

• Remove auto-consolidation: Auto-consolidation silently ran after bulk "Extract Now" and compressed hundreds of memories into 8 bullets — with no confirmation, no preview, and no undo. This has been removed entirely. Consolidation is now manual-only via the Consolidate button, which shows a before/after preview and supports undo.

Documentation

• Backup warning: Added a "Before You Start" section to the README advising users to back up their Data Bank files before using the extension.
• Consolidation docs updated: Clarified that consolidation is manual-only, added backup advice, noted that the undo is session-only.

1.2.0

New Features

• NVIDIA provider: Added NVIDIA as a built-in provider. NVIDIA's API doesn't support CORS, so requests are automatically routed through SillyTavern's server proxy — no extra setup needed.
• Reasoning/thinking model support: Models that use reasoning tokens (e.g., GLM-4.7 on NVIDIA) put output in reasoning_content instead of content. CharMemory now reads this automatically. Verbose logging shows [reasoning: N chars] when reasoning tokens are used. Some providers may support disabling reasoning via the system prompt field — see README for details.
• API key reveal/hide toggle: Eye icon button next to the API key field to show/hide the key. Auto-hides after 10 seconds for security.
• Connect/Test Model flow: Explicit Connect button fetches the model list with inline status feedback. Test Model button verifies the selected model responds correctly, showing model name, response time, and whether it followed the test instruction.
• Verbose API response logging: When verbose mode is enabled, the Activity Log shows HTTP status codes, finish reasons, token usage, and reasoning content length for all API calls.
• Character card in extraction prompt: The character card is now included as a bounded reference section so the LLM knows what baseline traits NOT to re-extract. This significantly reduces card-trait leakage.

Improvements

• Default chunk size reduced: "Messages per LLM call" default changed from 50 to 20. Testing showed 50 caused timeouts with some providers, and 20 produces good results for most chat styles.
• Response length slider max increased: From 2000 to 4000 tokens, to accommodate reasoning/thinking models that need budget for both reasoning and output.
• Default to Dedicated API: Extraction source now defaults to "Dedicated API" instead of "Main LLM". Dedicated API produces better memories because the extraction prompt isn't polluted by chat context.
• Clearer UI labels: "API Provider" renamed to "Dedicated API", "LLM Provider" renamed to "LLM Used for Extraction".
• Extraction prompt refinements: Reduced card-trait leakage, meta-narration, and play-by-play through iterative prompt testing across multiple models.

Bug Fixes

• Detect proxy-forwarded API errors: SillyTavern's proxy returns HTTP 200 even for upstream errors, wrapping them in { error: { message } }. CharMemory now checks the response body for errors instead of relying solely on HTTP status.
• Add CSRF token to proxy requests: Proxy requests now include the CSRF token required by SillyTavern's server.
• Clear stale test status on provider switch: Test result text no longer persists when switching between providers.
• Fix extraction counter display: The progress counter in the stats bar now updates immediately when the interval slider changes.
• Fix stuck "Testing..." button: The Test button no longer gets stuck in the "Testing..." state on errors.
• Guard model fetch when API key is missing: No longer fires an API request with a blank key when the provider is selected before entering credentials.

Documentation

• Quick Start guide: New 4-step minimal setup guide at the top of the README.
• Full Setup Guide: Detailed walkthrough with screenshots, separated from the Quick Start.
• Reasoning model guidance: New section explaining how thinking models work and how to configure response length.
• "Why Local Vectorization Is Fine": New section explaining why local embedding models are adequate for CharMemory.
• Extraction tuning guidance: Advice on adjusting chunk size based on chat style, expectations for batch-extracting long chats.
• NVIDIA provider notes: Documentation for the transparent proxy routing.

1.1.0

New Features

• Injected memories diagnostics: The diagnostics panel now shows which individual memory bullets were retrieved and injected by Vector Storage during the last generation. Shows bullet count and full text.
• Character lorebooks: Diagnostics shows a static list of lorebooks bound to the current character, including entry counts and trigger keys for each entry.
• Vectorization source/model: Vectorization status now displays the configured embedding source and model (e.g., "Yes (42 chunks) via transformers/nomic-embed-text").

Bug Fixes

• Fix vectorization status always showing "No": The vector API call was missing the source and model parameters, so it couldn't find the correct vector directory. Now reads these from extension_settings.vectors.
• Separate static and runtime lorebook info: World Info section split into "Character Lorebooks" (always shows bound books) and "Activated Entries — Last Generation" (shows what fired at runtime).

1.0.1

Security

• Fix XSS in consolidation preview: Memory bullet text, chat IDs, and dates in the consolidation Before/After preview are now escaped with escapeHtml() to prevent script injection from crafted memory content.