agent-memory-systems

agent-memory-systems covers the architecture of memory for intelligent agents, framing memory as the cornerstone that keeps interactions from starting from zero every time. It addresses the design problem of how to store and, more importantly, retrieve information: the skill's key insight is that memory quality equals retrieval quality, not storage quantity, so chunking, embedding, and retrieval strategies determine whether an agent remembers or forgets. Because the field uses inconsistent terminology, the skill adopts the CoALA cognitive architecture framework, distinguishing semantic memory (facts), episodic memory (experiences), and procedural memory (how-to knowledge).

It lays out guiding principles such as chunking for retrieval rather than storage, decaying old memories, preferring background memory formation over real-time, and testing retrieval accuracy before production. It surveys concrete tooling with guidance on when to use each: memory frameworks including LangMem (LangGraph agents, semantic/episodic/procedural types), MemGPT/Letta (OS-style hierarchical memory tiers with automatic paging), and Mem0 (user personalization); vector stores including Pinecone (managed, billions of vectors), Qdrant (Rust-based, best filtering), Weaviate (hybrid search, knowledge-graph features), ChromaDB (prototyping, ~20ms p50 at 100K vectors), and pgvector (good for under 1M vectors on PostgreSQL); and embedding models such as OpenAI text-embedding-3-large (3072 dims, $0.13/1M tokens), text-embedding-3-small (1536 dims, $0.02/1M tokens), nomic-embed-text-v1.5 (768 dims), and all-MiniLM-L6-v2 (384 dims, lowest latency). It includes code patterns for the three memory types, a LangMem implementation with semantic upsert plus episodic and procedural adds, runtime retrieval that assembles a user profile with relevant past experiences and skills, and a vector-store decision matrix comparing scale, managed options, filtering, hybrid search, cost, and latency, with Pinecone, Qdrant, and ChromaDB examples.

It targets agent and LLM engineers designing memory systems, and it scopes out adjacent responsibilities: vector-database operations belong to a data engineer, RAG pipeline architecture to an llm-architect, embedding-model selection to an ml-engineer, and knowledge-graph design to a knowledge-engineer. The source is vibeship-spawner-skills (Apache 2.0) and it is marked risk: safe.