Compaction

Compaction replaces an older transcript prefix with a durable summary while preserving recent turns.

Configure automatic compaction

import {
  AgentBuilder,
  createSummaryMemoryCompactor,
} from '@anvia/core'

const agent = new AgentBuilder('support', model)
  .memory(memoryStore, {
    savePolicy: 'turn',
    compaction: {
      maxMessages: 40,
      keepRecentUserTurns: 4,
      compactor: createSummaryMemoryCompactor(summaryModel, {
        maxTokens: 1024,
      }),
    },
  })
  .build()

When stored history plus the incoming prompt exceeds maxMessages, Anvia summarizes older messages and retains the configured number of complete user-led turns.

Choose the threshold

Start maxMessages around half of the message budget you want in the main prompt. Retaining three to six recent user turns usually keeps the live conversation intact.

Compaction adds a summarization model call before the main completion. Its tokens contribute to total run usage, so choose summaryModel deliberately for cost, quality, and data handling.

Storage requirements

The Prisma, Drizzle, Postgres, and SQLite adapters support atomic compaction. A custom store must expose MemoryCompactionStore and replace the transcript prefix only when its revision still matches.

Conflicting concurrent writes are retried once by default. Exhausted conflicts raise MemoryCompactionConflictError; failed or empty summaries raise MemoryCompactionError before the main model call.

Compaction is conversation housekeeping, not an audit archive. Keep run events and traces in their dedicated stores.