agent-memory-compactor

Background service for semantic summarization and compaction of long-term agent memories to maintain context efficiency.

Overview

As AI agents interact with users over long periods, their "memory" (history, context) can grow large and redundant. agent-memory-compactor solves this by:

1. Semantic Analysis: Classifying memory importance using LLMs.
2. Temporal Clustering: Grouping related memories within time windows.
3. Semantic Compaction: Using LLMs to summarize clusters of low-importance memories into high-density facts.
4. Critical Preservation: Keeping high-importance memories in their raw form for precision.

Features

• Production-Ready: Written in TypeScript with full type safety.
• Zo Integration: Built specifically for the Zo Computer ecosystem, using the internal /zo/ask API.
• Configurable: Adjustable thresholds for importance and compaction frequency.
• Daemon Mode: Can run as a background service to continuously maintain efficiency.

Installation

git clone https://github.com/Retsumdk/agent-memory-compactor.git
cd agent-memory-compactor
bun install

Usage

Add a new memory

bun run src/index.ts --add "User mentioned they like dark mode." --tags "ui,preference"

Run compaction

bun run src/index.ts --compact

List memories

bun run src/index.ts --list

Run as a daemon

bun run src/index.ts daemon --interval 60

Configuration

The service uses a config.json file:

{
"threshold": 8,
"maxEntriesPerCluster": 20,
"compactionIntervalDays": 7,
"storagePath": "./data/memories.json"
}

Architecture

CLI (index.ts) -> Compactor (compactor.ts) -> MemoryStore (store.ts)
-> ZoClient (client.ts) -> Zo API

License

MIT License

---

Built by Retsumdk