RASPUTIN Memory — Experimental Conversational Memory System
RASPUTIN is an archived research codebase exploring long-term conversational memory for LLM agents, partitioning retrieval across four fact-type lanes and reranking with a cross-encoder, published for its architectural findings rather than as a production system.
RASPUTIN maintains a persistent store of memories extracted from conversation and exposes them through a search API and an MCP server, splitting retrieval across four typed lanes - windows, facts, entities and events - before a global cross-encoder reranking step, rather than searching one undifferentiated pool. The project explored per-lane budget calibration and reciprocal rank fusion as alternatives to simple top-k truncation, evaluated against a 1,540-question benchmark.
The author explicitly stopped development after the project beat one target benchmark for a time but was overtaken by a competing system, judging the remaining gap too costly in engineering time and token spend to close, and archived the project with its architecture and final benchmark documented for others to build on.
It is written in Python and released under the MIT license; the repository accepts no further issues or pull requests.