Lessons in Evidence, Consensus & Memory

What diverse source retrieval, linguistic analysis and governed AI memory taught me about making systems that preserve evidence, context and uncertainty.

Type

Resource

Published

2026-08-23

Last updated

2026-08-27

Reading time

7 min

When I started building Lucid, I assumed analysing individual news articles would be one of the harder problems.

It was. But another problem proved more interesting.

You could retrieve ten highly relevant articles about an event and learn surprisingly little. Six might repeat the same report. Two might quote the same announcement. Another might summarise those six. Only one might contain genuinely independent reporting.

The search worked. The intelligence did not.

That distinction led us into research on diverse retrieval, source independence, publisher behaviour, emotional framing and, eventually, long-term AI memory.

They seem like separate problems. Increasingly, I think they are variations of the same one:

How can a machine interpret information without losing track of what was actually observed?

A few papers have been particularly useful in thinking about that question.

Relevant is not the same as useful

Search systems are generally good at finding material relevant to a query.

The problem begins once relevance has already been established.

If the next nine results repeat the first, what did retrieving them accomplish?

NEWSCOPE, introduced in Uncovering the Bigger Picture: Comprehensive Event Understanding Via Diverse News Retrieval, approaches this by separating relevance from information coverage. It first retrieves relevant material, then favours results that contribute something meaningfully different.

Consider six articles:

1. Reuters reports that a company will cut 2,000 jobs.
2. A newspaper republishes Reuters.
3. Another site rewrites Reuters.
4. A financial publication explains why the cuts are happening.
5. A local outlet interviews affected workers.
6. The company publishes its restructuring announcement.

All are relevant.

Only some expand our understanding.

That changed how I think about retrieval. The useful question is not only:

Which sources are most relevant?

It is also:

What does this source add that we do not already know?

Sometimes the answer is another publisher. Sometimes it is a primary document, a local perspective or genuinely independent reporting.

Sometimes the next article adds nothing.

More sources do not necessarily mean more evidence

Ten articles are not necessarily ten confirmations.

A Reuters report may be syndicated across several publications, repeated by blogs and then cited again elsewhere. One observation can quickly begin to resemble consensus.

Consider:

Reuters reports X.ABC independently reports X.The company confirms X.Five publications repeat Reuters.

There are eight articles, but not eight independent observations.

This matters because AI systems increasingly use large collections of public information to decide what is established enough to summarise, remember or act upon.

Publisher credibility helps, but it cannot solve the problem alone. A respected publication can repeat someone else's reporting. A lesser-known outlet can conduct original work. A company statement can be authoritative about what was announced while remaining partial about what it means.

There is no single “source quality” number that resolves all of this.

Publisher context is not article truth

Lucid has maintained publisher-level intelligence for a long time.

It is useful to know how a publication behaves over time, which subjects it covers and how its framing changes.

But there is an obvious trap.

A publisher with a historical political tendency does not make every article published under that domain equally biased.

The article still has to be analysed as an article.

That is why I increasingly think publisher characteristics should sit around the evidence, not replace analysis of the evidence itself.

Sentiment misses the interesting part

Positive, negative and neutral sentiment are useful classifications. They are also blunt.

Consider three headlines describing the same inflation result:

Inflation falls to 3.1%.

Inflation remains painfully above target.

Inflation cools faster than expected.

The event is essentially identical. The framing is not.

Fear and anger may both look “negative” to a sentiment model. Trust and anticipation may both look “positive”. Yet they communicate very different things.

Writing about COVID-19 vaccines, by Alfonso Semeraro, Salvatore Vilella, Giancarlo Ruffo and Massimo Stella, helped sharpen this distinction. The researchers analysed 5,745 articles and examined how emotional framing shifted across publications, vaccines and stages of the pandemic.

What interested me was not simply whether coverage was positive or negative.

It was how different publications emotionally framed the same underlying event.

That thinking influenced a concept we call an Emotional Fingerprint.

The point is not to assign a permanent emotional profile to a publisher. It is to analyse articles first, then look for patterns across a subject, period or body of coverage.

A publisher may frame one topic through fear and another through optimism. Its framing may change after an event.

That change can itself be intelligence.

Framing and factuality are different dimensions

Emotionally loaded writing is not necessarily false.

Neutral writing is not necessarily accurate.

A politically slanted article can contain correct facts. A reputable publisher can make mistakes. A poor source can occasionally publish something true.

These dimensions should remain separate.

Five publications may agree that a company cut 2,000 jobs while disagreeing completely about what the decision means.

That distinction has always interested me:

What holds up across the coverage, and where does interpretation begin to diverge?

Flattening those layers together makes the system simpler.

It also makes it less useful.

Memory has the same provenance problem

Long-term AI memory initially seemed like an entirely different technical problem.

It is not.

Suppose I tell an AI:

We decided not to launch in New Zealand this year.

The system might store:

Company is not launching in New Zealand.

Useful, perhaps. But something has disappeared.

Who said it? Was it inferred from a document? Did a model derive it? And what happened to “this year”?

Once a compressed fact becomes detached from its origin, it can acquire more certainty than the original information deserved.

That is why Eywa: Provenance-Grounded Long-Term Memory for AI Agents, by Resham Joshi, caught my attention.

Its central principle is wonderfully concise:

Evidence before belief.

Eywa preserves underlying evidence before deriving more convenient memory representations from it. It also separates evidence retrieval, memory and answer generation rather than hiding all three inside one opaque process.

A useful memory system should therefore answer more than:

What do we know?

It should also answer:

Why do we think we know it?

Memory needs time as well as provenance

Suppose Lucid remembers:

Customer uses Vendor A.

Six months later, the customer migrates to Vendor B.

The first statement has not necessarily become false. It may simply have become historical.

Zep: A Temporal Knowledge Graph Architecture for Agent Memory explores this problem through temporal knowledge and evolving relationships.

The principle matters more to me than the particular database architecture:

This is wrong and this used to be true are different states.

For an intelligence system increasingly concerned with answering what changed?, that distinction is fundamental.

Shared memory introduces governance

Personal memory is hard enough.

Organisational memory adds different questions.

Who supplied the information? Who can access it? Was it verified? When does it expire? Can it influence an answer? Can it justify an action?

These are not merely retrieval problems. They are governance problems.

Semantic Modelling of Organizational Knowledge as a Basis for Enterprise Data Governance 4.0, by Miguel AP Oliveira and colleagues, explores organisational knowledge through semantics, provenance, lineage and lifecycle.

It reinforced a principle we had already started encountering in the product:

Remembering more is not the goal. Remembering with enough context to use information responsibly is.

Transformation should not create authority

Yedidel Louck's Securing LLM-Agent Long-Term Memory Against Poisoning examines a particularly uncomfortable failure mode.

An agent reads something untrusted.

It summarises it.

A trusted system later repeats that summary.

The new version is stored again.

Eventually, the original source becomes difficult to see.

The information has travelled, but it has not necessarily become more trustworthy.

Louck's broader principle is one I think applies well beyond memory security:

Transformation should not create authority.

A model expressing something more clearly does not make it more true.

A summary of an unreliable source remains derived from an unreliable source.

Five repetitions of one observation do not become five independent observations.

And this is where news consensus and agent memory unexpectedly meet.

They are both provenance problems.

Knowing something does not mean interrupting someone

Persistent memory creates another temptation.

If an AI knows enough about you, why not proactively tell you anything it thinks you might care about?

Because that becomes unbearable very quickly.

PASK: Toward Intent-Aware Proactive Agents with Long-Term Memory, by Zhifei Xie and colleagues, explores proactive systems through demand detection and persistent memory.

The useful distinction is:

What does the system know?

versus:

Does what is happening deserve the user's attention?

Those are different questions.

Always-on intelligence only works if the system learns when not to interrupt.

Otherwise it has not solved information overload.

It has automated it.

There probably is no perfect memory architecture

It is easy to read enough AI-memory research and conclude that every system needs graphs, semantic extraction, multiple retrieval strategies, consolidation and a grand ontology.

Then you build something impressive that performs worse than the simpler system it replaced.

Are We Ready For An Agent-Native Memory System?, by Wei Zhou and colleagues, is a useful corrective. The authors evaluate memory through representation, extraction, retrieval and maintenance.

Their conclusion is inconvenient in the best way: no single architecture wins everywhere.

More sophisticated structures can impose substantial cost without improving every workload. Repeated abstraction can also discard information that later proves important.

Sometimes the original mess contains exactly what you need.

Likewise, Memory-Orchestrated Semantic System, by Serge Lacasse, Jérémie Hatier and Alex Baker, approaches memory through structured relational storage and reproducible retrieval.

I like that partly because it is a reminder that not every AI problem demands exotic infrastructure.