30 ways to stop your AI agent from being that forgetful friend

What it does

This repository is a curriculum in notebook form. Each of the 30 techniques lives in a runnable Jupyter notebook, grouped into six families: short-term context tricks (sliding windows, summarization), long-term storage (vector stores, cross-session memory), cognitive architectures (working memory, hierarchical reflection), retrieval patterns, production frameworks (Mem0, Letta, Zep, Graphiti), and evaluation benchmarks including LoCoMo. The author also provides a decision tree and a comparison matrix to help you pick based on constraints like token cost, persistence needs, and retrieval style.

The interesting bit

Most memory tutorials stop at “use a vector database.” This one treats memory as a systems-design problem with a taxonomy: short-term vs. long-term vs. cognitive architecture vs. production deployment. The inclusion of newer entrants like Graphiti (episodic-to-semantic extraction) and Mem0 alongside established patterns like MemGPT’s self-editing memory makes it a snapshot of a field that is actively fragmenting into specialized infrastructure.

Key highlights

• 30 runnable notebooks, each technique isolated and executable
• Covers six families from conversation buffers to production deployment patterns
• Includes frameworks: Mem0, Letta (MemGPT), Zep, Graphiti
• Evaluation coverage: LoCoMo benchmarks and comparison matrices
• Decision tree and learning paths for technique selection
• Apache 2.0 licensed, Python 3.10+

Caveats

• The README is heavy on promotional material for the author’s book, newsletter, and social channels; the actual content depth per technique is hard to gauge without opening notebooks
• Some framework references (MemOS, Memori) are mentioned as “landscape” trends but not included in the 30 techniques
• “2026” dating in the README appears to be a forward-looking framing device rather than actual future content

Verdict

Worth bookmarking if you’re building agents and need to compare memory approaches empirically rather than architecturally. Skip if you want a single integrated framework — this is explicitly a survey, not a product.