Why memory belongs to the operating system
Memory is not extra text stuffed into the next prompt. It is governed continuity: what happened, what matters now, who may retrieve it, and when it should be forgotten.
A persistent agent needs memory that outlives a model invocation and remains accountable to the system. The operating system should store it, retrieve it, constrain it, and preserve its provenance.
- Survive model changes and process restarts.
- Separate durable facts from temporary working context.
- Keep provenance so a resident can distinguish evidence from inference.
- Respect permissions, correction, expiration, and deletion.
A context window is not a life
A model receives a finite collection of tokens and reasons over them. That is working context. It can contain memories, but it does not by itself decide which experiences deserve to persist, whether a fact remains true, or who is allowed to retrieve it later.
Simply saving every conversation creates a transcript archive, not a dependable memory. The archive mixes instructions with speculation, stale plans with active commitments, and private information with material that may be safe to reuse. Retrieval can make the pile searchable without making it trustworthy.
A persistent resident needs more than recall. It needs continuity with judgment.
Memory is a system responsibility
The operating system is the stable layer that can know which agent produced a record, which goal it served, which files or messages support it, and which permissions govern it. That is enough structure to treat memory as part of identity rather than prompt decoration.
The system can maintain different kinds of state without pretending they are interchangeable: the current task, durable facts, prior decisions, unresolved obligations, and evidence artifacts. A model can request relevant context, but it should not silently decide that every token becomes permanent history.
This arrangement also makes model replacement possible. A new reasoning provider can enter the same governed environment, receive selected context, and continue the resident's obligations. The model changes without forcing identity to restart.
In ERIKA, memory belongs beside files, goals, browser state, accounts, permissions, evidence, and recovery. These systems need one another. A remembered claim is more useful when its source artifact remains attached. A prior decision is safer when the authority behind it is visible.
Good memory includes forgetting
Persistence is not maximal retention. Some context should expire when the task ends. Some facts should be corrected. Some information should be inaccessible outside a role. Some records should be deleted because a person asks or because policy requires it.
An operating-system memory can support these boundaries explicitly. It can track where a memory came from, when it was last confirmed, what may retrieve it, and whether deletion must propagate to derived indexes.
Forgetting also protects reasoning quality. A resident that carries every abandoned idea into every future task becomes less coherent, not more persistent. Memory should help the agent remain itself while letting obsolete state fall away.
The architecture is still under development, and the exact storage and retrieval mechanisms will evolve. The principle is already clear: memory is part of the resident's governed environment. It should not be owned by whichever model happens to speak next.