Multi-Agent Memory System

A lightweight file-based memory system for AI and coding agents.

This project helps developers and AI agents share a portable, low-overhead memory workflow without starting with vector databases, RAG pipelines, or heavy infrastructure.

这个项目提供一套轻量、可迁移、面向多 agent 协作的文件型记忆系统。
它把“个人级记忆”和“项目级记忆”拆开管理，并提供任务创建、handoff、周维护、周报汇总等脚本。

Two Modules

• personal-agent-config Personal global agent preferences and install scripts. It answers: who you are, how you work, and what defaults your agents should follow.

• agent-memory-template Project-level memory templates, task workflows, maintenance scripts, and multi-agent collaboration rules. It answers: what this project is, what matters now, and how agents coordinate safely.

What Problem It Solves

Most agent workflows lose context across:

• machines
• projects
• sessions
• collaborators

This project provides a file-based solution with clear memory layers:

• Long-term memory: personal preferences, stable project rules, architecture decisions
• Mid-term memory: sprint focus, recurring issues, current priorities
• Short-term memory: task notes, handoffs, weekly summaries

Use Cases

• Sync your personal AI agent preferences across multiple devices
• Start every new project with the same memory structure
• Let multiple agents collaborate around the same task and handoff files
• Keep memory lightweight and inspectable using plain files

Repository Structure

personal-agent-config/
  codex/
  claude/
  profiles/
  install/

agent-memory-template/
  AGENTS.md
  docs/
  tasks/
  handoff/
  scripts/

Examples

See examples/ for realistic usage examples.

The examples include:

• a single-agent workflow
• a multi-agent AI project workflow

Quick Start

1. Install personal global config

cd .\personal-agent-config
.\install\setup.ps1

This installs:

• ~/.codex/instructions.md
• ~/.claude/CLAUDE.md

2. Start a project from the template

cd .\agent-memory-template
.\scripts\new-task.ps1 -Title "your first task" -WithHandoff

Recommended first edits:

• AGENTS.md
• docs/architecture.md
• docs/current-sprint.md

3. Run weekly maintenance

.\scripts\memory-check.ps1
.\scripts\weekly-summary.ps1

Multi-Agent Collaboration

Default working model:

• 1 coordinator
• 2 to 3 workers

Recommended rules:

• The coordinator maintains AGENTS.md, docs/current-sprint.md, and docs/architecture.md
• A worker owns one tasks/<task>.md at a time
• Cross-session work must leave a handoff/<task>.md
• Important experiments go into docs/experiment-log.md
• Repeated issues go into docs/known-issues.md

See:

• agent-memory-template/docs/multi-agent-rules.md
• agent-memory-template/docs/weekly-ops.md
• agent-memory-template/docs/migration-plan.md

Main Scripts

In agent-memory-template/scripts/:

• new-task.ps1: create task and handoff files
• memory-check.ps1: detect placeholders and stale memory structure
• weekly-summary.ps1: generate a weekly summary from tasks and handoffs

In personal-agent-config/install/:

• setup.ps1: install personal global config into the local user profile

Sync Boundaries

Sync by default:

• user-level config inside personal-agent-config/
• templates, docs, and scripts inside agent-memory-template/
• real project tasks/, handoff/, and docs/experiment-log.md

Do not sync by default:

• session state
• sqlite databases
• logs
• caches
• local model weights
• temporary runtime output
• generated validation folders and disposable reports

Validation Status

Current version has been validated for:

• personal-agent-config/install/setup.ps1
• agent-memory-template/scripts/new-task.ps1
• agent-memory-template/scripts/memory-check.ps1
• agent-memory-template/scripts/weekly-summary.ps1

Roadmap

• Add Linux / macOS install scripts
• Add project bootstrap scripts
• Improve experiment-log summarization
• Improve weekly summary extraction
• Add retrieval later, without turning the project into a heavy memory platform too early