Graph RAG vs sliding windows: How are you managing long-horizon agent memory?

Hey everyone! 👋

As autonomous AI agents shift from simple single-shot tasks to running multi-hour background jobs, standard context window management is hitting a serious wall.

Passing an entire git log or endless message history into large context models leads to rapid context rot, lost instructions in middle tokens, and soaring inference costs. Simple sliding window truncation loses critical historical decisions, while basic vector search RAG often misses structural code dependencies across nested directories.

Because of this, modern developer tooling is moving toward Continuous Context Compression and Knowledge Graphs. Instead of dumping raw chat logs or file dumps into the context, tools build dynamic entity graphs and AST representations in local vector stores, passing only distilled state snapshots to the LLM on each execution loop.

For engineers and makers building multi-agent systems—what memory architecture has given you the best balance between accuracy and token efficiency: Knowledge Graph RAG, recursive state summarization, or custom KV cache compression?