An operational context architecture for agents serving one entity over long horizons: memory as continuity rather than search — a temporal compression pyramid with immutable facts, and a delta recording protocol. Applied in Mojarrb’s Co-Founder mode.
Existing memory architectures for autonomous agents share one design assumption: that memory is a search problem. That assumption fails under the operational conditions of long-running agents serving a specific entity over extended time horizons.
An agent managing procurement, legal and HR for a business across eighteen months does not need to search general knowledge. It needs to know what this specific entity did, what its connected systems reported, and what decisions were made — structured, compressed and current at every session start. The same holds for a clinical agent serving a patient.
A five-layer temporal compression pyramid that preserves critical parameters through every compression cycle; a delta recording protocol that captures operational events automatically from both agent executions and connected systems; and a hierarchical traversal protocol that locates original artifacts without injecting file content into the session.
Applied in Co-Founder mode: it knows the business’s numbers and history, so an owner can put any decision to it without re-explaining the business from zero each time.
DOI: 10.2139/ssrn.6817840
ORCID: orcid.org/0009-0005-6512-3401