Edition History

This is not a changelog. It records what the second edition argued, what it established, and — just as deliberately — what it did not change.

What the first edition established

The first edition began from a strict definition and refused to widen it:

Memory is when retained past experience changes what the system does now.

Everything else was treated as mechanism, and each mechanism was required to repair a failure that a simpler system had actually demonstrated. Strong hybrid retrieval was kept as a serious rival rather than demoted to a strawman. Raw history stayed canonical beneath everything derived from it. Historical truth and current truth were held apart. Retrieval causality was kept distinct from evidential support. Relevance was held to be different from authority, and behavioural change was required as evidence rather than assumed from the fact that something had been retrieved.

The earned architecture is small: canonical history, strong retrieval, optional derived views, state and evidence resolution, an explicit work frame, frame-conditioned candidates, bounded context, a reader, observable behaviour, and evaluation — with provenance, temporal validity, unresolved work, context trace, and policy versions cutting across all of it.

The book also named what it refused. Lateral inhibition did not earn itself. The Nexus did not become an accuracy layer. Consolidation, forgetting, reinforcement, and procedures were kept out of the architecture because they resemble familiar taxonomy, not because evidence dismissed them.

The question the first edition left open

The first edition assumed that retained experience stays in the form it was stored. Everything downstream — retrieved, derived, resolved, framed, admitted — preserves that assumption without ever examining it.

At some point a system stops treating an observation as evidence about the past and starts treating it as an instruction about the present. A remark becomes a rule. A workaround becomes a policy. A decision becomes a constraint that future work must satisfy.

The first edition had no position on that transition, and its ladder could not reach it. All six questions begin after the representation has been chosen. None of them asks what kind of persistence a piece of experience deserves in the first place.

What triggered the second edition

In October 2026 Kevin Liao published an essay arguing that agents do not need memory but documentation, and that the memory-plugin ecosystem is solving the wrong problem at the category level. A Hacker News thread followed.

The essay did not prove the first edition wrong, and this edition does not present it as having done so. Its claims are an author’s argument with no measurement behind them, and the essay’s own companion document makes an unsourced universal claim about an entire market that the first edition’s Chapter 12 results directly contradict.

What it did was apply pressure to one specific sentence the first edition had asserted and never tested:

Derived state may be worth persisting for cost, but persistence does not automatically earn epistemic authority.

Documentation is the most persuasive persisted representation available. When an agent writes a document and a person approves it, it looks exactly like knowledge. If that sentence was wrong, this is where it would show.

What was added

A new Chapter 18, “When Memory Becomes Documentation.” It states the challenge at full strength before answering it, reads the working system the essay shipped rather than the essay alone, and treats the Hacker News thread as community discussion rather than evidence.

A named operation. Promotion: the controlled transformation of retained experience into a durable representation intended to govern a broader class of future situations. The chapter’s central finding is that this operation changes authority rather than merely representation, which is why it cannot be treated as another retrieval step.

Failure classes for promotion. Over-promotion, under-promotion, wrong scope, self-authored authority, stale promotion, and lost provenance, each kept as a distinguishable pair in the book’s existing style rather than collapsed A named operation. Promotion: the controlled transformation of retained experience into a durable representation intended to govern a broader class of future situations. The chapter advances the hypothesis that this operation changes authority rather than merely representation, which is why it cannot be treated as another retrieval step. That hypothesis is unrun.

A measurement instrument for promotion. Seven metrics, scored separately, with abstention scored in both directions so that a system which always promotes and one which never promotes both lose.

Two designed experiments. MEM-DOC-01 compares six persistence strategies over a project history that mutates after durable knowledge exists. PROMOTE-01 measures whether a system can decide what kind of persistence an experience deserves without overgeneralising it.

What changed in the architecture

Nothing. That is the accurate statement, and it should not be softened.

MEM-DOC-01 and PROMOTE-01 were designed and have not been run. The environment available for this edition had no model-provider credentials and no inference adapter, and the benchmark harness supports scripted policies only. Under the book’s standing rule — no mechanism is promoted without a demonstrated failure — promotion therefore enters the architecture as EXPERIMENTAL / DEFERRED, alongside Chapter 15’s growth policies and Chapter 16’s learning boundary.

What was run is narrower and is stated as a result about the measuring instrument only: a check that the promotion scorers separate the failure classes they are meant to separate. Twelve frozen cases, scripted policies, no model inference, reproducibility address promo-instrument-20261004T143000Z. The oracle reaches the ceiling on all seven metrics; over-promotion and under-promotion fail on different metrics in the predicted directions.

That run produced no claim about agents. It is included because a measurement that cannot tell its own failure modes apart is not yet an instrument.

The synthesis chapter’s diagram is consequently unchanged. A deferred mechanism drawn into the measured path would misrepresent what the experiments established.

What did not change

Everything the first edition earned. Specifically:

• Storage is not retrieval, and retrieval is not memory.
• Strong retrieval is a serious baseline, not a strawman.
• Raw history remains canonical beneath every derived layer.
• Derived state is a hypothesis about history, rebuildable where possible.
• Historical truth and current truth can both be correct at different standpoints.
• Memory is durable; context is selected.
• Derivation lineage is not evidential support.
• Relevance does not grant authority.
• Behavioural evidence is the only evidence that matters for the book’s own definition.
• There is no composite score.

The capstone’s 103-trace replay is untouched, and the existing platform architecture was not rewritten.

What remains open

• Whether a promoted artifact reaches behaviour more often than an unpromoted one. Nothing measures this.
• Whether catalog-mediated discovery beats similarity ranking on the same corpus.
• How often promotion is correct, and whether any model can distinguish a one-off review comment from a standing preference.
• Whether personal-scope knowledge belongs in project documentation at all. This is the case where over-promotion is a privacy failure rather than a maintainability problem, and it remains the least understood.
• Whether the reconstruction reading of Naur’s argument — that memory preserves evidence from which understanding can be rebuilt, rather than understanding itself — survives contact with a working system.
• Whether promotion is learning, which Chapter 16 deliberately declined to claim.

Why this merited a second edition

The honest answer is narrower than the marketing answer would be.

The second edition earned no new mechanism and moved no existing one. Its contribution is a question the first edition could not ask, an operation the first edition could not name, and a measurement apparatus for that operation that is ready to be run when a reader can.

That is a real contribution, but it is worth being clear about what it is not. It is not an advance in the architecture. It is the identification of a boundary the architecture does not yet cover, together with the discipline to say so instead of filling the gap with a plausible mechanism.

The first edition asked how the past becomes useful in the present.

The second adds a harder question, which it does not answer: when does experience stop being merely recalled and start being what the system believes the project is — and who is authorised to make that decision on the project’s behalf?

Asking it properly required adding almost nothing to the architecture. That is the more interesting result.