archon-memory-core

archon-memory-core is an open-source memory layer designed for agents that need to manage and resolve contradictory information over time. Unlike standard retrieval systems that may retain conflicting facts and return results based solely on embedding similarity, this tool actively consolidates memory, ensuring that newer, superseding facts replace outdated or contradicted ones. The platform addresses the challenge of memory degradation in agent systems, where accumulating contradictions can lead to unreliable outputs.

0 license. 10+ on Linux and macOS (both arm64 and x86_64 architectures). Its core functionality includes embedding-based retrieval with type-weighted salience, nightly consolidation passes that cluster memory by entity co-occurrence, and compression using a local language model. Supersede-aware consolidation ensures that when a new fact contradicts an old one, the outdated information is archived rather than allowed to persist and interfere with retrieval. The system also differentiates between types of stored data, such as credentials (which never decay) and session notes (which expire), and ranks information based on persistence class and explicit priority, not just embedding similarity.

archon-memory-core has been benchmarked with a simulated ninety-day horizon, injecting daily contradictory facts and confusable distractors to test its ability to maintain accurate, consolidated memory. 2% for retrieval-only memory and 0% for context-dump or word-overlap baselines. The benchmark uses a realistic token budget with FIFO eviction and does not rely on artificially unlimited context windows.

This memory layer is intended for developers building agent systems that require robust handling of evolving and potentially conflicting information. Its open-source nature and detailed benchmarking position it as a specialized solution for long-term memory management in agent architectures.