Spreading activation retrieves relevant items by labeling source nodes with activation weights, then iteratively propagating activation through linked nodes in an associative network. section 14.3
In LLM memory systems, a concept graph organizes past interactions so retrieval can follow relationships between extracted concepts rather than relying only on direct similarity. section 14.3 section 14.4
A practical retrieval flow combines spreading activation with other memory mechanisms:
- Extract concepts and generate embeddings from user–model interactions. section 14.3
- Represent interactions as a graph whose linked concepts support relevance-based retrieval. section 14.3
- Combine retrieved long-term interactions with recent short-term history when generating a response. section 14.3
- Reinforce frequently accessed memories and allow unused memories to decay. section 14.3 section 14.4
Memoripy memory retrieval
Memoripy builds a NetworkX-based concept graph from prompts and responses, then uses spreading activation to retrieve relevant interactions alongside recent conversation history. section 14.3
