Memory Extraction

The MemoryExtractionService (apps/api/src/arena/memory-extraction.service.ts) uses Claude to extract structured memories from conversation transcripts. It handles all three operations/surfaces.

Extraction Model

All extraction calls use claude-haiku-4-5 for fast, cost-effective processing.

Methods

`extractAgentMemories(agentName, transcriptText)`

Extracts 3–8 individual memories for a specific agent from an arena voice transcript.

Categories: decision, position, task, open_question, learning

Used by: Arena session end handler

`extractFromChat(agentName, chatHistory)`

Extracts 3–6 memories from a 1:1 text chat conversation.

Categories: decision, position, task, open_question, learning, preference

Used by: Agent chat message handler

`extractFromTeamChat(teamName, agentNames, chatHistory)`

Extracts both per-agent and team-level memories from a group chat.

Returns:

{
  agentMemories: Record<string, ExtractedMemory[]>,
  teamMemories: ExtractedMemory[]
}

Used by: Team chat message handler

`extractSessionBrief(transcriptText, agentNames)`

Extracts a structured session brief from an arena transcript.

Returns:

{
  decisions: { text: string; confidence: string }[],
  actionItems: { owner: string; task: string; deadline?: string }[],
  unresolved: { topic: string; positions: string[] }[],
  nextSessionQuestions: string[]
}

`consolidateMemories(memories)`

Consolidates a cluster of related memories into a single comprehensive memory. Used by the weekly operations/cron-jobs|consolidation cron.

Input: Array of { content, memoryType } objects

Output: Single { content, memoryType } with 2–3 sentence summary

Memory Object Structure

All extraction methods return arrays of:

interface ExtractedMemory {
  category: 'decision' | 'position' | 'task' | 'open_question' | 'learning' | 'preference';
  content: string;      // 1-2 sentence concise statement
  confidence: 'high' | 'medium' | 'low';
}

JSON Parsing

The service uses defensive parsing with parseJsonArray() and parseJsonObject():

Regex extracts the first [...] or {...} from the LLM response
Falls back to empty array or default object on parse failure
Logged warnings for debugging

operations/surfaces — Where extraction is triggered
pipeline/lifecycle — What happens after extraction
pipeline/scoring — How extracted memories are later retrieved
operations/cron-jobs — Consolidation that uses consolidateMemories()

Memory Extraction

On this page