The field named its pains separately
Each is what one part of an un-closed loop looks like from outside. Different symptoms; one shared failure surface; one grammar of repair — status · provenance · supersession · scope · closure — carried on the objects that move through the loop.
Seven open problems
Questions the program is instrumented to study and does not claim to have answered — every branch feeds the same spine: operational closure at scale.
The depth law of drift
Does context degradation compound multiplicatively with delegation depth while cost merely adds — and what is the exponent? If provenance-carrying packets flatten the curve, by how much, at what packing cost?
The research program is built on four core positions.
Closure means something precise — and it hasn't been defined
Completion is a property of a task; closure is a property of an operation. An operation is closed when what was known, who authorized it, what was done, what resulted, and what memory now says are one traceable object.
Sovereign and portable at the same time
The context packet refuses to trade one for the other: a bounded, governed, portable assembly of context prepared for a specific recipient, task, and moment — the right context, at the right time. The corpus stays home; the projection travels.
Integrity lives in the coupling
Every response is elicited from a unique state space — prompt, memory, tools, environment. The answer was never a property of the model alone. So drift, bloat, and sycophancy are attacked by engineering what enters the loop: epistemic agentic memory, with status, provenance, supersession, stakes, and validity carried on the objects themselves.
Humans stay accountable; the loop stays traceable
Agents propose; promotion into durable state is a governed, observable, human-accountable act. Every claim traces to the packet that scoped the work that produced it. Judgment is routed, not automated away.
Together these define the category the series builds toward: governed agentic operations — and the layer that enables them, governed agentic infrastructure.
The series, in reading order
Published here first, then Zenodo, then arXiv.
Open Problems at the Agentic Memory Wall
A Research Agenda for Governed Agentic Operations
States the four commitments, re-describes the field's symptoms as failures of closure, and poses the seven problems above — each falsifiably, mapped to the paper and benchmark tier designed to test it.
Megan Anderson · AI ARMY, INC. · aiarmy.co · 2026
Integrity Is a Property of the Coupling & Trust Is Symbiotic with the Environment in Agentic Operations
Defines the category and the closure claim.
Context Packet Specification v0.4.0
Defines the primitive that travels — the right context, at the right time.
Governance Starts at Capture: An Epistemic Memory Architecture for Governed Agentic Operations
Defines the epistemic agentic memory the loop writes into.
Epistemic Context Packets as an Inference-Economics Primitive: Valid Handoff and the Agentic Memory Wall
Prices the loop — valid handoff as the economic unit; pre-registers the program's falsification conditions.
Memory as a constellation; shaping the information space
Memory objects are stored as stars or nodes; provenance links are the bridges between them, creating clusters or constellations over time. Unrelated objects or ungoverned objects remain isolated but accessible until promoted to a cluster. A packet is constructed from the information field of constellations for the right context at the right time. The memory stays home; the packet travels.
Reference the research program
Governance is not overhead — it is how agentic systems preserve the conditions of valid action
The harness, task sets, and scoring schema are being prepared for release.