Causal Memory

•
110
•

Provides AI agents with a long-term memory system that records and reasons about decision-outcome causal relationships, including inhibitory connections.

Causal Memory is an advanced agent memory system designed to prevent AI agents from repeatedly making the same mistakes by preserving crucial causal information. Unlike conventional memory solutions susceptible to context window compaction, this system stores `decision → outcome` relationships in an external SQLite database, ensuring survival through numerous interactions. It features a hippocampus-inspired engine that uses typed spreading activation (both excitatory and inhibitory), Hebbian learning, and Q-value dynamics to allow agents to not only recall facts but also understand *why* past actions worked or failed, and *what would happen if* they acted differently.

주요 기능

01Implements immutable SWR consolidation for long-term memory maintenance, including LTP/LTD/GC and meta-edge pattern mining.
02Preserves causal information outside the agent's context window, immune to context compaction.
0347 GitHub stars
04Supports forward simulation ('intervention_query') for predicting outcomes and 'counterfactual_query' for comparing alternatives.
05Models both excitatory and inhibitory causal relationships (e.g., 'caused', 'prevented') for nuanced reasoning.
06Utilizes a hippocampus-inspired engine with typed spreading activation, Hebbian co-occurrence, and Q-value dynamics.

사용 사례

01Analyzing the root causes of past failures through multi-hop backward causal chain tracing.
02Preventing AI agents from repeatedly making past mistakes by retaining decision-outcome causal links.
03Predicting potential outcomes and warnings before an AI agent takes an action.