mnemos: Biomimetic Memory Architectures for Large Language Model Agents

Current memory systems for LLM agents rely on appendonly vector stores that grow monotonically, fail to filter noise at ingestion, and retrieve by semantic similarity alone—ignoring the cognitive state of the user and the associative structure of stored knowledge. We present mnemos, an open-source Python library that implements five neuroscience-inspired memory mechanisms as composable modules: (1) a surprisal gate based on predictive coding theory that filters low-information inputs at write time; (2) mutable RAG that reconsolidates memories on retrieval, solving the stale-fact problem; (3) an affective router that blends emotional-state similarity into retrieval scoring; (4) a sleep daemon that consolidates episodic interactions into semantic abstractions; and (5) spreading activation over an associative memory graph. We provide ablation studies demonstrating each module’s contribution, showing that the surprisal gate reduces stored noise by 40% at the default threshold, affective routing achieves perfect statecongruent retrieval in controlled settings, and spreading activation with 20% decay reaches 4/4 nodes in a concept chain versus 1/4 at 90% decay. mnemos is MCP-native, supports three storage backends (inmemory, SQLite, Qdrant, neo4j), and includes a memory safety firewall. All code is MIT-licensed and available athttps://github.com/anthony-maio/mnemos.