Show HN: MemLedger – AI agent memory you can trust (github.com)

MemLedger, a new AI memory framework showcased on HN, promises to solve the "black box" issue of agent memory by providing full transparency on the provenance of each memory. While traditional frameworks allow AI agents to remember user preferences across sessions, they often don't reveal how or why certain information is retained or lost. MemLedger addresses this by offering a clear audit trail for every memory, making it easy for developers to track the origin and decisions behind each stored fact. This transparency enhances debugging capabilities and trust in AI interactions. Significantly, MemLedger employs an append-only event ledger structure that mirrors human cognition with three memory layers: instinct (core facts), episodic (long-term memories), and working (current session data). Unlike other systems that freeze memories at the time of writing, MemLedger allows for continuous improvement by regenerating memories with advanced models over time. This system not only reduces memory costs by filtering non-essential data but also implements anti-poisoning measures by requiring confirmation of new facts across multiple sessions before they become permanent. As a lightweight solution that runs locally without vendor lock-in and supports various AI models, MemLedger could represent a pivotal shift in how the AI/ML community approaches memory management and usability in applications.