claude-brain 0.15.2 → 0.17.0

package/src/memory/episodic/summarizer.ts

@@ -1,179 +1,179 @@
-/**
- * Extractive Summarizer
- * Generates summaries by scoring and selecting key sentences
- */
-
-import type { Episode, EpisodeSummary } from './types'
-
-interface ScoredSentence {
-  text: string
-  score: number
-  position: number
-}
-
-export class ExtractiveSummarizer {
-  private briefSize: number
-  private detailedSize: number
-
-  constructor(briefSize: number = 2, detailedSize: number = 8) {
-    this.briefSize = briefSize
-    this.detailedSize = detailedSize
-  }
-
-  summarize(episode: Episode): EpisodeSummary {
-    const allText = episode.messages.map(m => m.content).join('\n')
-    const sentences = this.splitSentences(allText)
-
-    if (sentences.length === 0) {
-      return {
-        brief: 'Empty episode',
-        detailed: 'No messages in this episode.',
-        key_topics: [],
-        decisions_made: [],
-        unresolved_questions: [],
-        entity_count: 0
-      }
-    }
-
-    // Calculate word frequencies for TF scoring
-    const wordFreqs = this.calculateWordFrequencies(allText)
-    const totalSentences = sentences.length
-
-    // Score each sentence
-    const scored: ScoredSentence[] = sentences.map((text, position) => ({
-      text,
-      position,
-      score: this.scoreSentence(text, position, totalSentences, wordFreqs)
-    }))
-
-    // Sort by score descending
-    const ranked = [...scored].sort((a, b) => b.score - a.score)
-
-    // Select top sentences for brief and detailed, then re-order by position
-    const briefSentences = ranked
-      .slice(0, this.briefSize)
-      .sort((a, b) => a.position - b.position)
-      .map(s => s.text)
-
-    const detailedSentences = ranked
-      .slice(0, this.detailedSize)
-      .sort((a, b) => a.position - b.position)
-      .map(s => s.text)
-
-    // Extract key topics from high-frequency words
-    const keyTopics = this.extractKeyTopics(wordFreqs)
-
-    // Find decisions (sentences with decision keywords)
-    const decisionKeywords = ['decided', 'chose', 'selected', 'will use', 'going with', 'recommend', 'should use']
-    const decisions = sentences.filter(s =>
-      decisionKeywords.some(k => s.toLowerCase().includes(k))
-    ).slice(0, 5)
-
-    // Find unresolved questions (last messages with ? that aren't answered)
-    const lastMessages = episode.messages.slice(-Math.ceil(episode.messages.length / 2))
-    const questions = lastMessages
-      .filter(m => m.role === 'user' && m.content.includes('?'))
-      .map(m => {
-        const qSentences = this.splitSentences(m.content)
-        return qSentences.filter(s => s.includes('?'))
-      })
-      .flat()
-      .slice(0, 3)
-
-    return {
-      brief: briefSentences.join(' '),
-      detailed: detailedSentences.join(' '),
-      key_topics: keyTopics,
-      decisions_made: decisions,
-      unresolved_questions: questions,
-      entity_count: keyTopics.length
-    }
-  }
-
-  private splitSentences(text: string): string[] {
-    // Split on sentence boundaries
-    const raw = text
-      .replace(/\n+/g, '. ')
-      .split(/(?<=[.!?])\s+/)
-      .map(s => s.trim())
-      .filter(s => s.length > 10)
-
-    return raw
-  }
-
-  private calculateWordFrequencies(text: string): Map<string, number> {
-    const stopWords = new Set([
-      'the', 'a', 'an', 'is', 'are', 'was', 'were', 'be', 'been', 'being',
-      'have', 'has', 'had', 'do', 'does', 'did', 'will', 'would', 'could',
-      'should', 'can', 'may', 'might', 'to', 'of', 'in', 'for', 'on',
-      'with', 'at', 'by', 'from', 'as', 'and', 'but', 'or', 'not', 'so',
-      'if', 'that', 'this', 'it', 'its', 'i', 'me', 'my', 'we', 'our',
-      'you', 'your', 'he', 'she', 'they', 'them', 'their', 'what', 'which',
-      'who', 'when', 'where', 'how', 'all', 'each', 'both', 'few', 'more',
-      'most', 'other', 'some', 'no', 'only', 'own', 'same', 'than', 'too',
-      'very', 'just', 'about', 'up', 'out', 'off'
-    ])
-
-    const freqs = new Map<string, number>()
-    const words = text.toLowerCase()
-      .replace(/[^a-z0-9\s-]/g, ' ')
-      .split(/\s+/)
-      .filter(w => w.length > 2 && !stopWords.has(w))
-
-    for (const word of words) {
-      freqs.set(word, (freqs.get(word) || 0) + 1)
-    }
-
-    return freqs
-  }
-
-  private scoreSentence(
-    sentence: string,
-    position: number,
-    totalSentences: number,
-    wordFreqs: Map<string, number>
-  ): number {
-    let score = 0
-
-    // 1. TF-based keyword score
-    const words = sentence.toLowerCase()
-      .replace(/[^a-z0-9\s-]/g, ' ')
-      .split(/\s+/)
-      .filter(w => w.length > 2)
-
-    if (words.length === 0) return 0
-
-    let keywordScore = 0
-    for (const word of words) {
-      keywordScore += wordFreqs.get(word) || 0
-    }
-    keywordScore /= words.length // normalize by sentence length
-    score += keywordScore * 0.4
-
-    // 2. Position bonus (first and last sentences score higher)
-    const posRatio = position / Math.max(totalSentences - 1, 1)
-    const positionScore = posRatio < 0.15 ? 1.0 : posRatio > 0.85 ? 0.8 : 0.3
-    score += positionScore * 0.25
-
-    // 3. Decision keyword boost
-    const decisionKeywords = ['decided', 'chose', 'selected', 'recommend', 'should', 'must', 'will use', 'going with', 'use', 'implement', 'adopt']
-    const lowerSentence = sentence.toLowerCase()
-    const hasDecision = decisionKeywords.some(k => lowerSentence.includes(k))
-    if (hasDecision) score += 0.2
-
-    // 4. Length penalty for very short or very long sentences
-    const optimalLength = 20 // words
-    const lengthRatio = Math.min(words.length, optimalLength * 2) / optimalLength
-    const lengthScore = lengthRatio <= 1 ? lengthRatio : 2 - lengthRatio
-    score += Math.max(0, lengthScore) * 0.15
-
-    return score
-  }
-
-  private extractKeyTopics(wordFreqs: Map<string, number>): string[] {
-    return Array.from(wordFreqs.entries())
-      .sort(([, a], [, b]) => b - a)
-      .slice(0, 10)
-      .map(([word]) => word)
-  }
-}
+/**
+ * Extractive Summarizer
+ * Generates summaries by scoring and selecting key sentences
+ */
+
+import type { Episode, EpisodeSummary } from './types'
+
+interface ScoredSentence {
+  text: string
+  score: number
+  position: number
+}
+
+export class ExtractiveSummarizer {
+  private briefSize: number
+  private detailedSize: number
+
+  constructor(briefSize: number = 2, detailedSize: number = 8) {
+    this.briefSize = briefSize
+    this.detailedSize = detailedSize
+  }
+
+  summarize(episode: Episode): EpisodeSummary {
+    const allText = episode.messages.map(m => m.content).join('\n')
+    const sentences = this.splitSentences(allText)
+
+    if (sentences.length === 0) {
+      return {
+        brief: 'Empty episode',
+        detailed: 'No messages in this episode.',
+        key_topics: [],
+        decisions_made: [],
+        unresolved_questions: [],
+        entity_count: 0
+      }
+    }
+
+    // Calculate word frequencies for TF scoring
+    const wordFreqs = this.calculateWordFrequencies(allText)
+    const totalSentences = sentences.length
+
+    // Score each sentence
+    const scored: ScoredSentence[] = sentences.map((text, position) => ({
+      text,
+      position,
+      score: this.scoreSentence(text, position, totalSentences, wordFreqs)
+    }))
+
+    // Sort by score descending
+    const ranked = [...scored].sort((a, b) => b.score - a.score)
+
+    // Select top sentences for brief and detailed, then re-order by position
+    const briefSentences = ranked
+      .slice(0, this.briefSize)
+      .sort((a, b) => a.position - b.position)
+      .map(s => s.text)
+
+    const detailedSentences = ranked
+      .slice(0, this.detailedSize)
+      .sort((a, b) => a.position - b.position)
+      .map(s => s.text)
+
+    // Extract key topics from high-frequency words
+    const keyTopics = this.extractKeyTopics(wordFreqs)
+
+    // Find decisions (sentences with decision keywords)
+    const decisionKeywords = ['decided', 'chose', 'selected', 'will use', 'going with', 'recommend', 'should use']
+    const decisions = sentences.filter(s =>
+      decisionKeywords.some(k => s.toLowerCase().includes(k))
+    ).slice(0, 5)
+
+    // Find unresolved questions (last messages with ? that aren't answered)
+    const lastMessages = episode.messages.slice(-Math.ceil(episode.messages.length / 2))
+    const questions = lastMessages
+      .filter(m => m.role === 'user' && m.content.includes('?'))
+      .map(m => {
+        const qSentences = this.splitSentences(m.content)
+        return qSentences.filter(s => s.includes('?'))
+      })
+      .flat()
+      .slice(0, 3)
+
+    return {
+      brief: briefSentences.join(' '),
+      detailed: detailedSentences.join(' '),
+      key_topics: keyTopics,
+      decisions_made: decisions,
+      unresolved_questions: questions,
+      entity_count: keyTopics.length
+    }
+  }
+
+  private splitSentences(text: string): string[] {
+    // Split on sentence boundaries
+    const raw = text
+      .replace(/\n+/g, '. ')
+      .split(/(?<=[.!?])\s+/)
+      .map(s => s.trim())
+      .filter(s => s.length > 10)
+
+    return raw
+  }
+
+  private calculateWordFrequencies(text: string): Map<string, number> {
+    const stopWords = new Set([
+      'the', 'a', 'an', 'is', 'are', 'was', 'were', 'be', 'been', 'being',
+      'have', 'has', 'had', 'do', 'does', 'did', 'will', 'would', 'could',
+      'should', 'can', 'may', 'might', 'to', 'of', 'in', 'for', 'on',
+      'with', 'at', 'by', 'from', 'as', 'and', 'but', 'or', 'not', 'so',
+      'if', 'that', 'this', 'it', 'its', 'i', 'me', 'my', 'we', 'our',
+      'you', 'your', 'he', 'she', 'they', 'them', 'their', 'what', 'which',
+      'who', 'when', 'where', 'how', 'all', 'each', 'both', 'few', 'more',
+      'most', 'other', 'some', 'no', 'only', 'own', 'same', 'than', 'too',
+      'very', 'just', 'about', 'up', 'out', 'off'
+    ])
+
+    const freqs = new Map<string, number>()
+    const words = text.toLowerCase()
+      .replace(/[^a-z0-9\s-]/g, ' ')
+      .split(/\s+/)
+      .filter(w => w.length > 2 && !stopWords.has(w))
+
+    for (const word of words) {
+      freqs.set(word, (freqs.get(word) || 0) + 1)
+    }
+
+    return freqs
+  }
+
+  private scoreSentence(
+    sentence: string,
+    position: number,
+    totalSentences: number,
+    wordFreqs: Map<string, number>
+  ): number {
+    let score = 0
+
+    // 1. TF-based keyword score
+    const words = sentence.toLowerCase()
+      .replace(/[^a-z0-9\s-]/g, ' ')
+      .split(/\s+/)
+      .filter(w => w.length > 2)
+
+    if (words.length === 0) return 0
+
+    let keywordScore = 0
+    for (const word of words) {
+      keywordScore += wordFreqs.get(word) || 0
+    }
+    keywordScore /= words.length // normalize by sentence length
+    score += keywordScore * 0.4
+
+    // 2. Position bonus (first and last sentences score higher)
+    const posRatio = position / Math.max(totalSentences - 1, 1)
+    const positionScore = posRatio < 0.15 ? 1.0 : posRatio > 0.85 ? 0.8 : 0.3
+    score += positionScore * 0.25
+
+    // 3. Decision keyword boost
+    const decisionKeywords = ['decided', 'chose', 'selected', 'recommend', 'should', 'must', 'will use', 'going with', 'use', 'implement', 'adopt']
+    const lowerSentence = sentence.toLowerCase()
+    const hasDecision = decisionKeywords.some(k => lowerSentence.includes(k))
+    if (hasDecision) score += 0.2
+
+    // 4. Length penalty for very short or very long sentences
+    const optimalLength = 20 // words
+    const lengthRatio = Math.min(words.length, optimalLength * 2) / optimalLength
+    const lengthScore = lengthRatio <= 1 ? lengthRatio : 2 - lengthRatio
+    score += Math.max(0, lengthScore) * 0.15
+
+    return score
+  }
+
+  private extractKeyTopics(wordFreqs: Map<string, number>): string[] {
+    return Array.from(wordFreqs.entries())
+      .sort(([, a], [, b]) => b - a)
+      .slice(0, 10)
+      .map(([word]) => word)
+  }
+}

package/src/memory/episodic/types.ts

@@ -1,52 +1,52 @@
-/**
- * Episodic Memory Types
- * Type definitions for the episodic memory system
- */
-
-export interface EpisodeMessage {
-  role: 'user' | 'assistant' | 'system'
-  content: string
-  timestamp: string
-  token_estimate?: number
-}
-
-export interface EpisodeSummary {
-  brief: string
-  detailed: string
-  key_topics: string[]
-  decisions_made: string[]
-  unresolved_questions: string[]
-  entity_count: number
-}
-
-export type SessionClassification = 'new_session' | 'continuation' | 'topic_shift'
-
-export interface Episode {
-  id: string
-  project?: string
-  status: 'active' | 'completed'
-  started_at: string
-  ended_at?: string
-  messages: EpisodeMessage[]
-  summary?: EpisodeSummary
-  related_decisions: string[]
-  related_patterns: string[]
-  related_corrections: string[]
-  token_count: number
-  message_count: number
-}
-
-export interface EpisodeMetadata {
-  episode_id: string
-  project: string
-  status: 'active' | 'completed'
-  started_at: string
-  ended_at: string
-  message_count: number
-  token_count: number
-  key_topics: string
-  brief_summary: string
-  related_decisions: string
-  related_patterns: string
-  related_corrections: string
-}
+/**
+ * Episodic Memory Types
+ * Type definitions for the episodic memory system
+ */
+
+export interface EpisodeMessage {
+  role: 'user' | 'assistant' | 'system'
+  content: string
+  timestamp: string
+  token_estimate?: number
+}
+
+export interface EpisodeSummary {
+  brief: string
+  detailed: string
+  key_topics: string[]
+  decisions_made: string[]
+  unresolved_questions: string[]
+  entity_count: number
+}
+
+export type SessionClassification = 'new_session' | 'continuation' | 'topic_shift'
+
+export interface Episode {
+  id: string
+  project?: string
+  status: 'active' | 'completed'
+  started_at: string
+  ended_at?: string
+  messages: EpisodeMessage[]
+  summary?: EpisodeSummary
+  related_decisions: string[]
+  related_patterns: string[]
+  related_corrections: string[]
+  token_count: number
+  message_count: number
+}
+
+export interface EpisodeMetadata {
+  episode_id: string
+  project: string
+  status: 'active' | 'completed'
+  started_at: string
+  ended_at: string
+  message_count: number
+  token_count: number
+  key_topics: string
+  brief_summary: string
+  related_decisions: string
+  related_patterns: string
+  related_corrections: string
+}