Skip to main content

2 posts tagged with "long-running-agents"

View All Tags

The Complete AgentOS Cognitive Memory Architecture

· 11 min read
AgentOS Team
Core Contributors

This is the complete technical reference for how AgentOS gives AI agents genuine memory — not RAG lookups, but a cognitive system grounded in published neuroscience research that encodes, decays, consolidates, archives, rehydrates, and perspective-encodes memories across hundreds of sessions.

We also cover how wilds.ai uses this system to give game NPCs and AI companions distinct minds that remember the same events differently.

Memory Archive: Lossless Cold Storage for Long-Running Agents

· 3 min read
AgentOS Team
Core Contributors

Temporal gist is one of the strongest features in AgentOS cognitive memory. Old, low-retrieval traces get compressed to 2-3 core assertions, mirroring how human episodic memory fades to semantic gist over time. But compression was destructive: the original verbatim content was overwritten with the summary, and only a SHA-256 hash remained for audit.

For agents that run across hundreds of sessions, this created a ceiling. After a few months, the agent's early memories were irreversibly compressed. Important details that happened to not be retrieved often enough were lost to the gist.

Today's release fixes this with IMemoryArchive and on-demand rehydration.