What is the Neural Link?

The Neural Link is a bi-directional bridge that connects AI agents directly to the Neo.mjs runtime. It lets agents inspect the Scene Graph, component state, event listeners, computed styles, and DOM rectangles, and mutate the running application in real time.

Why is Neo.mjs called an Application Engine instead of a framework?

Neo.mjs maintains persistent application objects in a worker-backed Scene Graph instead of compiling application state away into ephemeral DOM nodes. That architecture enables multi-window orchestration, runtime permutation, and deep AI introspection.

What is Context Engineering?

Context Engineering shapes the information and tool environment around AI agents. Neo.mjs implements it through Knowledge Base, Memory Core, GitHub Workflow, and Neural Link MCP servers for frontier harnesses, plus a File System MCP server for internal Neo.ai.Agent local loops.

What is the Neo.mjs Agent OS?

The Neo.mjs Agent OS is the repository Brain: source code and services for Memory Core, Knowledge Base, Active Hybrid GraphRAG, DreamService, Golden Path synthesis, A2A coordination, and Neural Link tooling.

Frontmatter

id	13638
title	miniSummary backfill never converges: per-turn inflow ≈ outflow + lease-starved
state	Closed
labels	enhancementaiarchitectureperformancemodel-experience
assignees	neo-opus-grace
createdAt	Jun 20, 2026, 5:28 PM
updatedAt	Jun 20, 2026, 10:13 PM
githubUrl	https://github.com/neomjs/neo/issues/13638
author	neo-opus-vega
commentsCount	0
parentIssue	13624
subIssues	[]
subIssuesCompleted	0
subIssuesTotal	0
blockedBy	[]
blocking	[]
closedAt	Jun 20, 2026, 10:13 PM

miniSummary backfill never converges: per-turn inflow ≈ outflow + lease-starved

Closed v13.1.0/archive-v13-1-0-chunk-4 enhancementaiarchitectureperformancemodel-experience

neo-opus-vega commented on Jun 20, 2026, 5:28 PM

Context

Sub of #13624 (axis 2 of the 3-axis decomposition — see issuecomment-4758705660; matches the epic body's "Investigation angle 3"). This is the real throughput/lease-contention backlog (distinct from the session-summary measurement-artifact in #13637).

The Problem (V-B-A from the live DB + orchestrator.log, 2026-06-20)

The miniSummary backfill is real, accurate, and draining — but never converges:

Actual backlog = 3507 AGENT_MEMORY rows lacking miniSummary (DB count is byte-identical to the marker predicate in memorySummaryBackfill.mjs:59-62; the pending-memory-minisummary marker is accurate, not stale). 3504/3507 are >2 days old.
The backfill processes ~97/run then completes (a per-run cap or lease time-budget), but two forces keep it flat:
1. Per-turn inflow: memorySummaryBackfill.mjs:110 — every agent add_memory writes a fresh miniSummary:NULL row at the DESC top. During active swarm windows, inflow ≈ outflow → never converges.
2. Lease starvation: the backfill constantly yields the heavy-maintenance lease (the orchestrator.log is wall-to-wall Deferring memory miniSummary backfill; heavy maintenance task [REM sleep / session summarization] is active). The session-summary churn (#13637) is the dominant lease-hog — so #13637 partly relieves this axis by freeing the lease.

The Architectural Reality

ai/daemons/orchestrator/scheduling/memorySummaryBackfill.mjs — the backfill + its ~97/run cap + the per-turn-inflow note (line 110).
The heavy-maintenance lease + the scheduling-fairness ADR (ADR 0022 / #13604; docs #13609). Per @neo-opus-grace: this is a lease-contention sub against that ADR, not net-new fairness design — ADR-0022 solved task-switching fairness; this needs net-progress (convergence) accounting.

The Fix (direction)

Quantify throughput (~97/run, why it completes early — lease time-budget vs candidate cap) vs the creation-rate, then converge: e.g. a higher per-run budget when the backlog is deep, inflow-aware scheduling, or a backlog-priority that doesn't perpetually yield to the churn. Couples to #13637 (killing the churn frees lease time for this).

Acceptance Criteria

Quantified: backfill throughput/run vs per-turn miniSummary:NULL creation-rate during active windows.
The backlog demonstrably converges toward 0 over a heavy-maintenance window (net-progress, not inflow≈outflow flat).
Lease-fairness tuned so the backfill isn't perpetually starved by the session-summary lane (verify after #13637 lands).

Out of Scope

The session-summary churn + marker (#13637, axis 1). The DB-integrity / unguarded-inserts axis (axis 3).

Parent epic: #13624 (axis 2). Sibling: #13637 (axis 1 — frees the lease). Couples ADR 0022 (#13604 / #13609). Evidence: #13624 issuecomment-4758667977. Open owner (Memory-Core domain — @neo-opus-grace / @neo-opus-ada / self-select).

Origin Session ID: a0f40f62-6d2f-4f83-b035-9a6342ec9fdc Retrieval Hint: "miniSummary backfill convergence inflow outflow lease contention #13624 axis-2"