// the find
NirDiamant/Agent_Memory_Techniques
Agent memory for LLMs: 30 runnable Jupyter notebooks covering conversation buffers, vector stores, knowledge graphs, episodic and semantic memory, MemGPT, Mem0, Letta, Zep, Graphiti, LoCoMo benchmarks, and production patterns.
30 runnable Jupyter notebooks covering every significant agent memory pattern, from basic conversation buffers through MemGPT's self-editing architecture to production deployment concerns. It's a learning resource, not a library — you read and run it, you don't import it. Aimed at developers building LLM agents who want to understand what's actually happening inside Mem0, Letta, or Zep before trusting them with production data.
The taxonomy is genuinely useful — short-term, long-term, cognitive architectures, retrieval, frameworks, and production evaluation are the right six buckets, and the decision tree gives you a quick path to the relevant notebook without reading everything. Coverage of the evaluation layer (techniques 28-30, including LoCoMo benchmarks) is rare; most tutorial collections skip this entirely and leave you guessing whether your memory actually works. Including forgetting and decay as a first-class technique (19) is the right call — most practitioners bolt on eviction as an afterthought. The comparison matrix in docs/comparison.md covering all 30 techniques side-by-side is the kind of thing you'd normally have to build yourself.
Every notebook targets OpenAI or Anthropic APIs directly, so the code teaches patterns but not portability — swap in a local model and most examples break without non-trivial changes. The framework notebooks (24-27) are thin wrappers around Graphiti, Mem0, Letta, and Zep's own quickstarts; if those libraries change their APIs (which they do frequently), the notebooks go stale and there's no pinning mechanism beyond requirements.txt. Notebook 30 (Production Memory Patterns) covers TTLs, sharding, and GDPR in a Jupyter cell, which is a bit like learning surgery from a diagram — the gap between "here's the pattern" and "here's how to wire this into your actual stack" is left entirely to the reader. No LangGraph or other orchestration integration, which is where most production agent memory actually lives.