Durable Agent Memory

A lightweight memory protocol for long-running AI agents.

AI agents lose context. Sessions end, agents change, experiments fail midway, and important decisions get buried in old conversations.

Durable Agent Memory gives agents a structured, evidence-backed way to preserve project state across context loss, interruptions, and handoffs.

What it does

• Preserves project state across sessions and agents
• Tracks decisions, failures, hypotheses, and next steps
• Links claims to evidence such as files, commits, run IDs, and artifacts
• Supports safe resumption by reconciling memory with the actual workspace
• Validates memory structure with a read-only Python validator

Operating modes

Mode	Purpose
Bootstrap	Create memory for an existing project
Update	Record new decisions, evidence, failures, or progress
Resume	Recover the live state after context loss or handoff
Audit	Check for stale claims, broken references, or memory drift

Example structure

project-memory/
├── README.md
├── USER_DIRECTIVES.md
├── CURRENT_STATE.md
├── DECISION_LOG.md
└── NEXT_CHECKPOINT.md

Memory stays compact and points back to the real project rather than trying to replace it.

The workspace is the source of truth. Memory is the index.

Validation

python scripts/validate_memory.py project-memory/ --strict

The validator checks memory structure, broken links, missing core files, unsafe paths, and other common issues.

Repository

├── SKILL.md
├── agents/
├── references/
└── scripts/
    └── validate_memory.py

Built for long-running development, research, and multi-agent workflows where losing the context should not mean losing the project.