sillytavern-character-memory/CLAUDE.md

CharMemory — SillyTavern Extension

Git / GitHub

This repo belongs to the bal-spec GitHub account. The .envrc file (via direnv) automatically sets GH_TOKEN for bal-spec when working in this directory — no manual gh auth switch needed.

What This Is

A SillyTavern extension that automatically extracts structured character memories from chat messages, stores them as markdown in the character's Data Bank, and relies on Vector Storage for retrieval at generation time.

File Structure

index.js        — All extension logic: extraction, consolidation, provider API calls, UI controllers, event handlers, modals (~7800 lines)
lib.js          — Pure utility functions imported by index.js at runtime and used by tests (parsing, serialization, formatting, stripping)
editor.js       — Shared memory block editor factory (createMemoryEditor) with state management and undo
settings.html   — Sidebar dashboard HTML (stats bar, extraction controls, tool launchers, activity, diagnostics)
style.css       — All styling (dashboard, modals, drawers, wizard, troubleshooter)
manifest.json   — ST extension manifest (version, loading order, author)
README.md       — User-facing documentation (getting started guide + technical reference, combined)
CHANGELOG.md    — Version history
images/         — Screenshots for documentation
test/           — Vitest test suites (unit, integration/snapshot, integration/live)

index.js is the main runtime module. lib.js is the canonical source for pure utility functions — index.js imports them via ES modules. Only serializeMemories() is kept local in index.js because it uses getFormatOptions() for runtime settings.

Key Architecture

Extraction Pipeline

1. Extension listens for CHARACTER_MESSAGE_RENDERED events
2. Counts messages against a configurable interval (default 20)
3. Collects unprocessed messages in chunks ("Messages per LLM call", default 50)
4. Strips non-diegetic content (code blocks, tables, HTML tags) from messages
5. Sends existing memories + recent messages + character card to the LLM with the extraction prompt
6. Parses <memory> blocks from response, appends to Data Bank file
7. Vector Storage handles vectorization and retrieval automatically

Provider System (v1.2.0)

Three extraction sources: EXTRACTION_SOURCE.MAIN_LLM, EXTRACTION_SOURCE.WEBLLM, EXTRACTION_SOURCE.PROVIDER.

The "Dedicated API" source uses PROVIDER_PRESETS — a registry of named presets (OpenAI, Anthropic, OpenRouter, Groq, DeepSeek, Mistral, xAI, NanoGPT, Ollama, Pollinations, Custom). Each preset defines: baseUrl, authStyle ('bearer' | 'x-api-key' | 'none'), modelsEndpoint ('standard' | 'custom' | 'none'), requiresApiKey, extraHeaders, defaultModel.

All providers except Anthropic use the OpenAI-compatible /chat/completions endpoint via generateOpenAICompatibleResponse(). Anthropic has its own adapter generateAnthropicResponse() that converts to the Messages API format.

Per-provider settings (API key, model, system prompt, custom URL) are stored in extension_settings.charMemory.providers[key].

LLM Dispatch

callLLM(userPrompt, maxTokens, defaultSystemPrompt) is the single dispatch point used by extraction, consolidation, and connection testing. It branches on extension_settings.charMemory.source.

Memory Format

<memory chat="main_chat_abc123" date="2024-01-15 14:30">
- Bullet point memories
- One block per extraction/encounter
</memory>

UI Layout (v2.0)

The sidebar (settings.html) is a single-view dashboard — no tabs. All complex UI is in center-screen modals and drawers built dynamically in index.js:

• Dashboard (sidebar): Stats bar, file info, extraction toggle, Extract Now, tool launcher buttons (Consolidate, Batch, Format), mini activity log, diagnostics summary
• Settings Modal (showSettingsModal()): Left-nav with sections — Connection, Extraction, Storage, Advanced. Uses cm_modal_* prefixed IDs to avoid conflicts with sidebar elements.
• Prompts Modal (showPromptsModal()): Full-width editor for extraction/consolidation prompts with version tracking and update banners
• Log Drawer (toggleLogDrawer()): Slide-out right-side drawer for the full activity log with verbose toggle and export
• Troubleshooter Modal (showTroubleshooter()): Health checks, Data Bank file browser, diagnostic report, reset/clear tools
• Setup Wizard (showSetupWizard()): 3-step first-run flow — LLM Connection, Vector Storage, Ready

Settings Storage

All settings live under extension_settings.charMemory. Key fields:

• source — extraction source enum
• selectedProvider — active provider preset key
• providers — per-provider settings objects
• extractionPrompt — customizable prompt template
• interval, cooldownMinutes, chunkSize, responseLength — extraction tuning
• perChat, fileName — storage options
• promptVersions — tracks which prompt versions the user has seen (for update notifications)

Conventions

• All UI element IDs are prefixed with charMemory_
• jQuery is used for DOM manipulation (ST convention)
• LOG_PREFIX = '[CharMemory]' for all console output
• logActivity() for user-visible activity log entries
• escapeHtml() for all user-generated content rendered as HTML
• Settings are saved via saveSettingsDebounced() (ST global) after any change
• The extraction prompt uses {{charName}}, {{charCard}}, {{existingMemories}}, {{recentMessages}} template variables

Important Patterns

• Never break the extraction prompt without testing — the prompt has been iteratively refined to reduce card-trait leakage, meta-narration, and play-by-play. Changes should be tested against multiple characters with varied content.
• Memory parsing is strict — only <memory> blocks with - bullet lines are recognized. Other formats will be silently ignored.
• Provider settings are isolated — switching providers preserves each provider's API key, model, and system prompt independently.
• NanoGPT has special model fetching — uses a custom endpoint with rich metadata (provider grouping, subscription/open-source/roleplay/reasoning filters). Other providers use the standard /models endpoint.
• CORS matters — this runs in a browser. Most providers support CORS but some (Ollama) require configuration (OLLAMA_ORIGINS=*).

Testing

Automated Tests

Vitest is the test framework. Three tiers:

npm test                # Unit tests (97 tests, ~200ms) — pure functions in lib.js
npm run test:snapshot   # Snapshot tests (6 tests) — extraction pipeline against 1000-message fixture
npm run test:live       # Live LLM tests (3 tests) — requires a running OpenAI-compatible server

Unit tests cover parsing, serialization, escaping, format detection, and the three extracted pipeline functions (stripNonDiegetic, formatChatMessages, substitutePromptTemplate).

Snapshot tests process real chat data from test/fixtures/flux-chat.jsonl through the pipeline and snapshot the output. Update snapshots with npm run test:snapshot -- --update after intentional changes.

Live LLM tests send extraction prompts to a real LLM and validate the response structure. Configured via env vars:

• TEST_LLM_URL — endpoint (default: http://127.0.0.1:1234/v1)
• TEST_LLM_MODEL — model name (default: auto-discover first available)
• TEST_LLM_KEY — API key for authenticated endpoints like OpenRouter (default: none)

Recommended local model: Gemma 2 9B or Qwen 2.5 7B. Avoid thinking models (Qwen3) — their <think> tags waste token budget.

lib.js as Single Source of Truth

lib.js is the canonical source for pure utility functions. index.js imports them via ES modules (import { ... } from './lib.js'). When modifying these functions, edit lib.js only — index.js picks up changes automatically. Exception: serializeMemories() in index.js is a separate implementation that uses getFormatOptions() for runtime format settings.

Manual Testing

For UI and integration testing that requires SillyTavern:

1. Install in SillyTavern's public/scripts/extensions/third-party/CharMemory (symlink or clone)
2. Restart SillyTavern
3. Test extraction with different providers
4. Check Activity Log (verbose mode via Log Drawer) for LLM prompts/responses
5. Open Troubleshooter to verify health checks and injected memories

Common Tasks

• Adding a new provider: Add entry to PROVIDER_PRESETS, no other changes needed if it's OpenAI-compatible with standard /models endpoint
• Modifying the extraction prompt: Edit defaultExtractionPrompt constant. Bump PROMPT_VERSIONS.extraction to trigger update notifications for existing users. Users can also customize via the Prompts modal, so changes to the default only affect new installations or users who click "Restore Default"
• Adding dashboard UI elements: Add HTML to settings.html, add event handler in the appropriate setup*Controls() function. Follow the charMemory_ ID prefix convention
• Adding modal UI elements: Build the HTML dynamically in the modal's show function (e.g., showSettingsModal()). Use cm_modal_* or cm_ts_* prefixed IDs to avoid conflicts with sidebar elements