When AI Proposed 412 Memory Edits, We Made Every Decision by Hand

Your company memory grows messy. Duplicate decisions pile up. Superseded checkpoints linger. Open questions multiply while their answers sit three entries down. You need cleanup, but who has time to audit thousands of memory units?

We built an AI gardener to propose cleanups. Then we learned why human oversight isn't optional—it's measurable precision.

The Memory Spine That Never Forgets

Our shared memory operates as an append-only spine. Every decision, checkpoint, and open question from humans and agents gets recorded. Nothing disappears. The spine grows, but it also accumulates near-duplicates and superseded units that clutter your operational view.

Enter the nightly gardener pass. This AI scans for near-duplicate or superseded memory units and proposes a simple action: mark the older unit as deprecated in favor of the newer one. Clean, surgical, reversible.

But proposals never auto-apply. They land in a human review tray. Accept or dismiss—every single one.

When 83% of AI Proposals Were Wrong

On July 20, we hand-scored the proposal tray. Results: 2 correct proposals out of 12 total. The AI was wrong 83% of the time.

Root cause analysis revealed the flaw. The gardener determined supersede direction by comparing random UUIDs—a self-join deduplication trick where b.id > a.id decides which unit stays active. This works for finding duplicates but fails catastrophically for directed claims about which memory unit actually supersedes another.

Roughly half the proposals pointed backwards, ready to deprecate the correct unit while preserving the outdated one.

The Fix: Follow the Work, Not the IDs

We rebuilt the direction logic around the authoritative lifecycle of work itself. Which ask closed? Which delegation merged? The gardener now derives supersede direction from actual workflow completion, not from UUIDs or timestamps that mean nothing for imported units.

The difference was immediate and measurable.

412 Decisions in Four Hours

On August 21, 2026, our founder drained the entire proposal tray live. Starting count: 92 pending proposals. The session ran about four hours, processing in waves: 92 → 43 → 10 → 0.

Final production database ledger:

• 262 proposals accepted
• 150 proposals dismissed
• 0 pending
• 89% acceptance rate in the final processing window

Nothing got bulk-applied. Every decision went through human review.

Why 89% Acceptance Still Requires Gates

That 89% acceptance rate represents measured precision for the AI proposer, not a trust setting. When the rate was 17%, the human gate caught the problem. At 89%, the gate remains fully active.

This isn't about building trust in AI. It's about publishing precision numbers you can defend.

The system includes a second-pass semantic contradiction judge that runs on your own LLM key with a hard 15-call limit per run. It only emits findings when the judge confirms a contradiction. Unsure pairs get dropped. Empty beats wrong, every time.

Agents Propose, Humans Merge

The principle holds: agents propose, humans merge. The merge gate's precision becomes a number you publish, not a feeling you have.

Your AI can scan thousands of memory units overnight and surface the cleanup candidates. But the decision to deprecate Unit A in favor of Unit B? That stays human. The precision measurement tells you whether your AI gardener is getting better or worse at proposing valid cleanups.

Why This Matters for Your Infrastructure

Memory management hits every team building with AI agents. Your agents generate decisions, checkpoints, and questions. Without active curation, your operational memory becomes a junk drawer.

But automated cleanup without human oversight creates a different problem: invisible errors that compound. When your AI deprecates the wrong memory units, you lose institutional knowledge without knowing it happened.

The measured approach gives you both benefits: AI-powered discovery of cleanup opportunities plus human-verified precision that you can track and improve over time.

Your company memory stays clean. Your oversight stays meaningful. Your precision numbers stay honest.

Building Your Own Memory Gates

Start with append-only memory storage. Add AI-powered cleanup proposals. Require human review for every merge decision. Measure and publish your acceptance rates.

When your AI proposes 412 edits to company memory, make sure a human decides every one. The precision number that results becomes your most important AI metric.

Human oversight isn't a bottleneck—it's your quality gate. And quality gates work best when they produce numbers you can defend.