Pack

Agent Memory System

Your agent forgets everything at midnight. This is the architecture that fixes it.

Any AI agent system that can read and write files (Claude, GPT, local LLMs), a text editor, and 15 minutes. Scripts are stdlib-only Python 3.10+ — no dependencies, no install step.

// what it is

Most agents are born fresh every session. They forget your preferences, repeat mistakes, and carry nothing across days or projects. Stuffing it all into the system prompt hits the token ceiling fast; a vector database gives you fuzzy recall without the crisp structure an agent needs to actually decide something.

The pack teaches the hybrid: canonical markdown for facts the agent can trust, vector retrieval for fuzzy search. Critically, it teaches lane separation — so a trading agent does not contaminate a writing agent and each lane stays coherent.

Three stdlib-only Python 3.10+ scripts ship with it, so the maintenance side is real code rather than advice: archival, lane validation, and a KPI audit.

// why it works

Two-layer architecture: canonical markdown plus vector retrieval
Lane separation, so agent memories cannot cross-contaminate
Auto-archival that keeps memory useful across months instead of bloating
KPI auditing — measure whether the memory system is actually helping