Overview
Agent memory has four separate problems.
First, single-session memory. During a long interaction, Hermes must remember both what the user said and what the agent already did. If the user requested direct language at the beginning, the preference should still apply later. If Hermes created a file or committed to a plan, it should not lose track of that work.
Second, knowledge updates. New facts can replace old facts. A current job should supersede a previous job. A new preference for direct scripts should take priority over an older preference for long explanations. Memory must not treat every stored statement as equally current.
Third, multi-session memory. Work is often distributed across Telegram, Slack, Hermes desktop, Codex, and other agents. A useful memory system must connect a plan created in one session with assets created in another and troubleshooting completed elsewhere.
Fourth, temporal reasoning. Memory must consider when a statement was true, what changed afterward, and which state applies now. A preference can change more than once. The system needs more than a list of facts; it needs an evolving interpretation.
Honcho addresses these problems by becoming a shared memory layer. Conversations from multiple tools pass through one system. It creates peer cards for people and agents, records relationships and inferred characteristics, and maintains different perspectives among participants. This is described as diachronic identity: identity understood as something that develops over time rather than as a permanent static record.
Honcho also performs a reasoning process over accumulated conversations. A component referred to as
Neurommancer turns raw chats into more useful memory by processing knowledge updates, logical implications, contradictions, relationships, and peer cards. This processing is compared to dreaming because the system revisits accumulated information and reorganizes it into a more coherent working memory.