Defense-grade memory for every AI touchpoint

Defense is not a “RAG vertical”—it is a trust vertical. Every system that uses AI—briefing copilots, ISR triage, logistics assistants, policy/ROE checkers, comms summarization, maintenance guides, and AI companions beside the warfighter—needs the same bar: outputs must be defensible, not merely fluent. Amnesis is not an AI that makes operational decisions. It is deterministic memory and reasoning infrastructure so that whatever the model says can be traced, explained, and audited against recorded inputs—not invented from weights or opaque retrieval.

It sits above existing feeds and tools, structuring ISR, HUMINT, SIGINT, sustainment, ROE, and communications into a time-aware memory with provenance and explicit relationships. The goal is trustworthy augmentation of human judgment across workflows, not a single chat-style Q&A feature—and not replacement of command authority.

Military demo (sign in) Better RAG demo flow

Why trust has to be architectural

A model alone cannot be “trusted” in the operational sense: you need what it was allowed to know, where each claim came from, and what was uncertain or contested at the time. Commodity patterns—prompt stuffing, similarity search without lineage, re-ranking without a record—fail for any high-stakes AI touchpoint, not only classic document Q&A. Amnesis keeps behavior anchored in stored, attributable facts and makes gaps and conflicts visible instead of smoothed over.

Warfighter & AI companion

More and more, a warfighter will talk to an AI—headset, handheld, vehicle crew station—under fatigue, noise, and time pressure. In that mode, fluency feels like truth. A hallucination is not a bad customer review: it can mean a wrong coordinate or restriction, a misstated ROE or no-strike rule, a garbled fragment of an order, or a plausible-but-false medical or evac detail. Any of those can produce death, serious injury, fratricide, or mission failure—and the failure mode is silent until it is too late, because the model sounded sure.

The companion must therefore be fed from governed memory: what was ingested, scoped, and versioned for that workspace and checkpoint—not from whatever the base model “believes.” When the record is thin or contradictory, the right behavior is visible uncertainty and provenance-backed excerpts, not a confident guess. “We don’t have that on record” must be allowed to beat plausible invention.

Ingest & memory model

Heterogeneous operational material—ISR products, HUMINT, SIGINT-derived content, sustainment data, ROE and no-strike constraints, geofencing policy, recorded communications and orders (where permitted)—lands as structured memory with provenance, timestamps / temporal validity where applicable, and explicit relations to other nodes (e.g. supports, contradicts, supersedes—per policy).

Non-negotiable: when sources disagree, information is not silently overwritten or merged into one narrative. Contradiction stays first-class in the memory plane, consistent with the client UX posture in AMN-SPEC-019.

Reflection, clusters, and visible uncertainty

Reflection and clustering organize material into stable regions (converging evidence under policy) and unstable regions (disagreement, thin evidence, or high drift). In operations, the absence of certainty is often more important than a single confident answer—so unstable regions must stay visible, not hidden behind a ranked “best answer.”

Decision support—not autonomous decision

For assessments such as route safety, ROE compliance, or resource tradeoffs, Amnesis does not replace commander’s intent. It presents:

1. Supporting evidence with provenance and recency
2. Opposing or conflicting evidence
3. What would change the call—e.g. validated HUMINT or new SIGINT

Example (illustrative): ISR from 48 hours prior may show no armed presence; HUMINT may claim irregular forces nearby. Both remain in memory with timestamps and lineage. The operator sees what is known, how fresh it is, and where it conflicts—not a forced merge into one tidy verdict.

Planning & field agents

Planning draws on verifiable inputs only; hard constraints and missing intelligence stay explicit rather than optimized away.

Field agents act as coordinators and verifiers against trusted memory—e.g. a move request checked against stored ROE, geofencing, and no-strike corpora with exact clause and source on violation. The model does not invent operational facts; it grounds in what was ingested and scoped.

Audit, checkpoints, and replay

Meaningful knowledge states should be checkpointable and replayable for after-action review, legal and policy accountability, and institutional learning. Investigators load the checkpoint that existed at decision time and see exactly what was known—including unresolved warnings versus “clear” assessments—without hindsight rewriting the record.

Trust model (summary)

Product expectations

For defense tenants, every surface that consumes the memory plane—not only “search over PDFs”—should respect the same rules: explicit workspace, checkpoint, and embedding / policy scope; provenance and contradiction visible in evidence flows; no operational claim without provenance where policy requires it; and RBAC / export packs per tenant policy. Gaps versus shipped UI belong in backlog, not mock data.

Non-goals

• Autonomous weapon engagement or lethal decisioning without human authority
• Replacing national or coalition C2—Amnesis is a memory plane, not a replacement COP
• Guaranteeing completeness of the battlefield—only integrity of what was recorded

Specification: Full narrative, cross-references (AMN-SPEC-001, 005, 006, 008, 009, 019, 020), and document control → docs/specs/AMN-SPEC-035_amnesis_military_defense_use_case_spec.md. Synthetic demo corpus (JTF-ALPHA) → docs/demo/AMN-DEMO-DATA-002_military_demo_workspace.md. Demos: use /military_demo.html first (signed-in military workspace UI). General evaluator UI (Test a Better RAG): /better_rag_demo.html. CLI bootstrap (docs/demo/MIL_BOOTSTRAP.md, seed script, curl) is second—for operators and future CI, not the primary evaluator path.

Military demo (sign in) Better RAG demo flow Security Governed recall layer More use cases