AI memory OS · agent memory · context memory
Memory for AI agents should decide what deserves to be remembered.
Entroly Memory OS is a local-first memory and context-control system for AI agents. It manages working, episodic and semantic memory under explicit capacity and token budgets, then combines recall with safety, consolidation, persistence, receipts and verification.
Why Entroly calls it a Memory OS
Working memory
Short-horizon task state with bounded growth and explicit token cost.
Episodic memory
Session/history memory with retention and recall behavior rather than permanent accumulation.
Semantic memory
Persistent promoted patterns protected from normal forgetting while global capacity remains bounded.
Budget-aware recall
Recall scores task relevance, retention, frequency, tier and importance, then selects by value per token.
Consolidation
Sleep-replay-style promotion lets important working memories become episodic and high-value episodic memories become semantic.
Safety + verification
Memory traffic can be screened for unsafe content and output can be checked against selected evidence.
Memory that improves without unbounded accumulation
Entroly models forgetting as a feature, not a failure. Weak memories can decay, frequently recalled high-retention memories can promote, and recall is constrained by the context budget. That creates an adaptive memory system without pretending that every stored item is permanently useful.
Multi-agent memory
The deeper Entroly stack includes redundant-message suppression, compliance checks, lesson sharing with feedback, and experimental privacy-preserving federation. These surfaces have different maturity levels; the shipped MemoryOS Python facade, CLI, persistence, safety scanning, deterministic stress benchmark, Context Receipts and WITNESS verification are the public production surface.