VoltMem: A Volatility-Aware Memory Layer for LLM Agents
VoltMem adds domain-specific volatility priors to LLM agent memory, resolving stale-vs-stable fact conflicts that flat memory layers like Mem0 miss.
VoltMem tackles a structural blind spot in agent memory systems: most treat a volatile fact like "I live in Berlin" with the same protection as a stable one like "I prefer concise answers." VoltMem instead assigns each memory a domain-specific volatility prior — personality, core preference, professional context, current project, mood — and uses a formal weight-and-threshold equation to decide how easily a fact should be overwritten versus defended.
In a head-to-head case study against the open-source Mem0 across three scenarios (location change, a one-off contradiction to a stable preference, and a mood shift), VoltMem kept current, non-conflicting facts in all three while Mem0 either kept stale data or adopted weak contradicting evidence. At retrieval time, VoltMem's down-ranking of stale volatile memories cuts stale-top-result rate to 0% versus 20% for cosine-only search, though on the LongMemEval-S benchmark it ties rather than beats cosine similarity.
The library ships as a pip-installable package with zero required core dependencies, optional embeddings support, and a LangChain integration, plus overridable domain volatility priors for custom use cases like CRM data. The author is upfront that these priors are currently hand-tuned, framing the release as an engineering artifact validated by continual-learning experiments rather than a benchmark-chasing leaderboard entry — useful signal for engineers building production agent memory.