neural-memory

Associative memory for AI agents uses a neural graph where memories connect via typed synapses and are recalled through spreading activation. This approach surfaces conceptually related information by traversing explicit relationships rather than relying on keyword matching.

Storing and recalling agent memories offline is achieved using a local neural graph with spreading activation, requiring zero LLM dependencies or external embeddings. The system uses Hebbian learning to strengthen frequently co-accessed memories and applies natural decay to idle ones.

Associative memory with spreading activation runs completely offline with zero LLM dependency, requiring no external embeddings or API calls. Memories persist in a local database, allowing independent brain instances to maintain long-term context, decisions, and errors without network connectivity.

Hebbian learning and memory decay handle contradictions by auto-detecting conflicting memories and prioritizing current knowledge. Frequently co-accessed memories strengthen over time through Hebbian learning, while idle memories fade naturally, ensuring the neural graph maintains relevance and consistency.

Limitations of using a local neural graph include relying entirely on explicit relationship traversal instead of semantic vector search, which may miss unconnected concepts. Memory relevance depends heavily on the accuracy of spreading activation and Hebbian learning parameters configured for the specific agent.