For DoD programs and primes
Local context and memory for AI-enabled work.
Perseus is an open-source software layer for current context, durable memory, and reviewable records around an approved model or workflow.
Bounded workshare. Perseus supplies the software layer. The program retains mission authority; the prime or integrator owns interfaces, verification, and accreditation.
Conflict handling / MemConflictSelf-run · 8 providers · 3,750 questions · measures reported separately
This is a company-run replication, not official benchmark inclusion, customer validation, production validation, or Government authorization.
Where it fits
One bounded role beside the mission system.
Perseus provides context, memory, and records. The program and its prime retain the surrounding mission responsibilities.
Start with the state of the work.
Perseus Context Engine builds a briefing from the sources the host chooses.
Carry decisions into the next task.
Perseus Vault keeps dated facts, corrections, and choices locally.
Leave a trail people can check.
Perseus Ledger records supplied events and references in a hash chain.
Evidence
Tested behavior. No certification claim.
Each number has a method, a condition, and a limit. Read the reports before using a result in an evaluation or acquisition document.
MemConflict / tested Perseus Vault configuration18 wrong
Across 3,750 questions in a self-run comparison of eight self-hosted configurations.
1,378 of 3,750 questions blank or abstaining (36.7%). This is a conflict-handling result, not production validation.
Read the eight-provider report ↗Deployment boundary
The program keeps the authority.
Run the components locally, on premises, in a private cloud, or in a disconnected environment. The program chooses the data, keys, network, and permissions.
See the security boundary →- Perseus
- Current context, local memory, and reviewable records.
- Program
- Model choice, data classification, identity, authorization, network, safety, and mission authority.
- Prime
- Hardware, system interfaces, verification, validation, accreditation, and field support.
Suggested evaluation
Measure the failure modes that matter.
A small, matched test tells the program more than a feature list.
Choose one workflow
Define the task, data boundary, approved model, and deployment environment.
Compare the paths
Run a stateless baseline, ordinary retrieval, and Vault with the same task set and token budget.
Review the result
Record correctness, stale assertions, abstention, latency, recovery, access controls, and reviewable records.