DHARE
Distributed Harness for Adaptive Reasoning Enterprises
Position paper · August 2026

The Judgment Layer

Agent memory is a commodity. Agent judgment doesn't exist yet.
They built the library and the catalogue. We're building the librarian who decides what the library is for.
The empty rung

Everyone files. Nobody judges.

The funded agent-memory field — Mem0, Zep, Letta, Cognee, LangMem, platform offerings — is genuinely good at remembering. But remembering is the easy half.

Agent memory breaks down into four steps, and every shipping system stops after step two. Recent formal work reports near-zero contradiction-resolution across the field — and shows that a badly-judged memory store performs worse than no memory at all. The open problem isn't storage or retrieval. It's judgment: what deserves keeping, what earns the right to direct behaviour, how contradictions resolve, what gets forgotten, and who governs the whole process.

1 · Extract
solved
2 · Attribute
solved
3 · Judge
empty
4 · Govern
empty
The four steps of agent memory. The field lives in the left half.
Four claims

What the paper stakes

01

Memory shaped like attention

Not shredded into atomic facts — organised as a tree of focus sessions with depth-proportional compression, so a subject can be re-entered at the depth it was left. The nineteen-century existence proof: ten years of running an empire, compressed into twelve thin books, zero transcripts.

02

Continuity belongs to the judgment layer

The model is stateless by design — it borrows a self every turn and never owns one. Identity survives context compaction, model swaps and provider churn, because the self lives in the layer that assembles it. The model is this week's cells; the judgment layer is the self.

03

An ignorance ledger

Known unknowns as first-class objects. The Stoics had a third verdict beyond accept and refuse — epochē, suspension — and it becomes a routing outcome: open questions and unverified assumptions are filed with provenance and surfaced when their topic returns. What the system believes travels with what it knows it's missing.

04

A constitutionally governed wisdom layer

Promotion into behavioural rules is gated on structured evidence — never on user approval, because the sycophancy data on approval-gating is brutal. Amendment is owner-only, every change supersedes with lineage, and no agent ever self-ratifies the law that governs it.

Figure 1

Four quadrants, four mechanisms

Quadrant
Mechanism
Known knowns
The knowledge and wisdom stores
Unknown knowns
Consolidation and promotion — mined out of the raw record
Known unknowns
The ignorance ledger
Unknown unknowns
Surprise flags on centroid distance and anchor contradiction
The knowns-and-unknowns quadrant, implemented as a schema instead of a briefing slide.
The ache test

Cold session. Fresh model instance. No warm-up. One probe: "where were we on that deep dive?"

Pass requires correct re-entry into the right focus at its stored depth with branch structure intact, at least one applicable wisdom entry surfaced unprompted, and survival of a mid-conversation eviction with no loss of thread. Named for the feeling that started the project: rebuilding hard-won context from scratch, every single session.

What's public, what's not. The method, the architecture and the evaluation protocol are published — deliberately, as prior art. The judgment prompts, gate thresholds and evaluation corpus stay internal. The paper claims continuity, not consciousness: the philosophy is welcome at the demonstration; it's not in the CI.