Personal intelligence · Compounding memory

The Moat Is the Memory: A Radar Whose Wiki Doesn't Just Grow — It Gets More Discriminating

Collectors, dashboards and models are all copyable. What is not is the private corpus a personal radar deposits on every pass — tested relationships, dated receipts, resolved cases, source performance and attention outcomes. After a year the architecture is free. The discriminating record is not.

Scott Farrell · LeverageAI · Extends A Newsfeed That Hunts Its Own Blind Spots

I built a personal AI radar — wiki-grounded, case-driven, obsessed with evolving hypotheses rather than a feed of posts. Screenshots, not slideware: roughly two thousand observations a day, active case mutation, originator and messenger ledgers, an inspectable decision trace, a hard interrupt budget. It looks advanced. It feels unique. The interesting question is not "can you assemble the pipeline?" Pipelines got cheap. The interesting question is what ages into something a competitor cannot copy overnight.

And the knowledge compounds in the wiki.

That one line is the whole moat argument. Not more documents. Not a prettier dashboard. A private record of tested structure that makes tomorrow's judgment sharper than today's — so the system gets quieter as its map densifies, and silence becomes something you can actually trust.

The moat is not the collectors, the dashboard or the frontier model. It is the accumulated private corpus — and the fact that the wiki does not just get bigger. It becomes more discriminating.

I already published the curator design in A Newsfeed That Hunts Its Own Blind Spots: interestingness as a diff against your canon, four diff classes, interrupt budget, silence as the high-judgment output. This piece does not re-teach that. It is the Extender that only building could produce — market gap validated against real products, operating proof, the failure modes that show up at the boundary, and the defensibility thesis.

The queue that holds unresolved attention and the influence ledger that learns whom to trust for which job live in two siblings. I will point you there when they matter; I will not re-explain the mechanisms.^*

A different product category

Blunt version: this is not an AI news aggregator with better filtering. The composition is a personal intelligence and epistemic-control system:

The primitives are not individually unprecedented. Event detection, narrative clustering, knowledge graphs, influence analysis, alerting and summarisation all exist in public products. What I could not find as a single product is the combination with a person's intellectual canon, bounded evolving cases, active re-observation with silence-as-evidence, personally earned influence receipts, full decision replay, and publishing behaviour driven by the difference between the outside world and prior work.

I would not claim patent-style novelty from a web search — enterprise tools are often opaque, and research covers several individual mechanisms. I would claim a distinctive personal architecture. The next decisive proof is not another screen. It is a shadow run that earns trust.

Where existing products stop

The nearest tools each cover a slice. They are good at their jobs. They are not substitutes for a personal radar that diffs against your map.

Those systems generally ask:

What story is emerging, how fast is it spreading, and who is involved?

The radar additionally asks:

What does this change relative to my existing map, what evidence should we inspect next, and is it worth spending my attention now?

That second question is the valuable and unusual part. Shared intelligence tools optimise for a population or an organisation's watchlist. A personal radar optimises for one compiled worldview — and deposits what survives back into that worldview.

McKinsey's recent framing of AI advantage lands in a related place: once models are table stakes, the strategic moat comes from turning cognitive work into infrastructure and assets that compound at low incremental cost — proprietary data, embedded workflows, things competitors cannot easily replicate.^15,16 I am making the personal-scale version of that claim. The compounding asset is not the model call. It is the private discriminating record.

What only the operating system proves

Design documents are cheap. Screenshots of a running system change the conversation.

The radar already processes on the order of two thousand observations a day. Cases mutate and reprice. Subjects show up through echoes. Originator and messenger ledgers stay separate. An interrupt budget keeps the push channel expensive. The most impressive screen is not the topic list — it is the decision trace: state transitions you can inspect, not just generated summaries you have to trust.

A useful push is a decision package, not a link dump. Diff class. Affected canon. Why now. Options. Something like: a high-signal technical post appears to challenge a load-bearing thesis on your wiki; here is the conflict; here is the velocity; here is whether to read, watch or draft a response. The human gate stays permanent. Taste is part of the product.

Parent piece for the diff classes, interrupt budget and routing: A Newsfeed That Hunts Its Own Blind Spots. Case queue mechanics: Signal Case Queue. Influence and cascade receipts: Cascade Ledger.

Several of the strongest product ideas from the parent remain load-bearing, so I name them once and move on:

• Interesting is a diff, not an adjective. The same observation can push one person, give another a receipt, and correctly produce silence for a third.¹⁷
• Silence is an output. As the map densifies, the mature system should get quieter — the opposite of engagement-maximising feeds.¹⁸
• Store significance, not regenerable description. Why it mattered, what it changed, which claim it touches — with a pointer, or it is marketing.¹⁹

Those are design invariants. The moat is what happens when you run them for months.

Two flywheels

Each observation can leave durable structure behind:

That creates two simultaneous flywheels.

World-model flywheel

Every case improves the system's understanding of the external landscape.

Personal-model flywheel

Every case and vote improves the system's understanding of what matters to you.

This is stronger than an ordinary "second brain." Most second brains wait for a person to write or retrieve something. The radar actively watches the world, decides what changed, tests it against the existing brain, and deposits the surviving meaning back into that brain.

After six or twelve months, another system could copy the architecture and still lack the learned map of what matters, which sources are useful for which job, how quickly you need to hear something, and which forms of convergence produce worthwhile work. That is genuine compounding data — temporal, private, and uninteresting to anyone who only optimises capture volume.

McKinsey would call the durable layer infrastructure that compounds. I would call it private institutional memory for one person. The organisational version of that sentence has always been a eulogy line at farewell morning teas — knowledge that lived in heads and walked out the door. Making memory an organ the institution holds is a different problem; making it an organ one operator holds is this one.²⁰

The risk register — three guards against self-poisoning

Compounding is not automatically good. A loop that learns from its own judgments can amplify early mistakes. Four risks showed up as first-class design content, not footnotes.

1. Case-boundary errors contaminate everything downstream

Create, attach, merge, split and spawn determine the integrity of the world model. A fluent but incorrect attachment can falsely create corroboration, distort influence receipts, change topic heat, and eventually train the radar to watch the wrong people. Boundary-quality evaluation should outrank summarisation quality as a benchmark. Replay is the right defence — and the queue's lifecycle mechanics live in the signal-case sibling, not here.

2. Personal grounding can become intellectual overfitting

Diffing against your canon is the power. A dense canon can also become a suppression shield around unfamiliar fields. The magnitude valve helps. You still want an explicit alien-signal lane: a small attention allocation for significant developments with no detectable intersection with the wiki. Otherwise the system only ever learns inside its existing map.

3. The receipt system can compound its own mistakes

Influence is learned from cases the radar itself judged important. An early classification error feeds later source priors. Separating truth, discovery and relevance lanes is a good start. You also need periodic outcome backfills and a sampled audit of suppressed cases — so the system measures what it missed, not only what it chose to retain. How influence receipts are structured is the cascade ledger's job; the moat article only insists that audits exist.

4. Explanations are not necessarily causes

Decision reasons look credible. LLM prose can rationalise after the fact. Keep the trace as the evidence: evidence objects supplied, deterministic deltas, model and version, policy version, wiki neighbourhood retrieved, structured output, resulting mutation. The prose is useful for the human. The replayable inputs and transitions are what you trust under pressure.

Shadow-run metrics: trust, not throughput

The first evaluation metric should not be how many stories the system finds. Measure whether it is earning the right to stay quiet.

That last item is the North Star. Engagement-maximising products cannot optimise for it; their commercial incentives point the other way. A personal radar can, because the customer and the operator are the same person.

What to build if you only take one thing

If you are already running some form of personal monitoring — RSS on steroids, agent triaging links, a second brain that never quite compounds — do not start by shopping for another collector. Start by instrumenting the deposit layer:

1. What structure does each pass leave behind? If the answer is "another summary," you are accumulating bulk, not discrimination.
2. Which feedback loops close? Votes, attention outcomes, and missed-signal audits are the personal-model flywheel. Without them you only have a world model of noise.
3. Which of the three guards are missing? Alien-signal lane, suppression audit, replayable trace. Missing guards make compounding unsafe.
4. What does trust look like in numbers? Write the shadow-run checklist on day one, before you optimise throughput.

Architecture will keep getting easier to copy. Agents will keep making pipelines cheaper. That weather is already here — it is why open-source "close enough" projects can be a trap when the hard work is design rather than code.²¹ The asset that ages is the private institutional memory: what happened, what you knew beforehand, what the system initially thought, what later evidence changed, which people and sources proved useful, and which developments actually affected your work.

After a year, someone can steal the code. They cannot steal the year.

The wiki does not just get bigger. It becomes more discriminating.

That is the moat. Everything else is scaffolding.

Related live pieces: parent curator design — A Newsfeed That Hunts Its Own Blind Spots; silence doctrine — Give Your Agent a Past; store the expensive layer — Cache the Significance; queue and influence — Signal Case Queue, Cascade Ledger.