The oldest architecture in computing

Dr. Werner Vogels · rss · 2026-09-08 source ↗ Reader 🖥️

Agent Memory Brain-Inspired Computing Compute & Energy Infrastructure Long-Running & Managed Agents Trust

The brain is the oldest and still the best computing architecture — its structured, decaying, background-consolidating memory is the template that both AI agent systems and AI-era professionals should deliberately imitate to stay effective and relevant.

practical · mixed — Framed as expository reflection (a case study of Kiro Crew's memory architecture read against neuroscience) but its real payload is practical advice for AI-era professionals and builders: study the brain's memory architecture as a design template, and personally 'evolve' by getting hands-on with new tools rather than treating them as threats. The narrative (Cortical Labs, Kiro Crew) carries the argument more than explicit premises do.

Structure

Part	Role
Opening reframe: 'Will AI take my job?' becomes 'Will AI make me obsolete? Not if you evolve.'	context
Personal practice: Vogels 'evolves' by doing real work with new tools (Kiro Crew) rather than forming opinions abstractly	evidence / method
Cortical Labs anecdote: lab-grown neurons play Pong on 20 watts vs. a GPU's 700 watts and hundreds of thousands of iterations	evidence (energy-efficiency data point)
Jeff Hawkins / 'A Thousand Brains': cortical columns each build local predictive models; memory consolidates during sleep; repetition strengthens, non-repetition weakens	theoretical grounding
Claim: 'the brain is the oldest architecture in computing' — older than Von Neumann, the abacus, mathematics, language	thesis statement
Kiro Crew case study: distinct memory types (markdown files, vector database, key-value index), background consolidation, decay/compression/pruning, human-editable memories, remote/autonomous operation	evidence — a concrete engineered system arriving at brain-like architecture independently
Trusted autonomy: memory lets agents earn the trust to work unattended	extension / consequence
Closing exhortation: 'the oldest architecture is still the best... those who study the brain will make the biggest breakthroughs'	conclusion / call to action

Key terms

evolve
Actively, hands-on adapting to new tools by doing real work with them long enough to form an honest, tested opinion — the individual-level analog of what he argues memory architectures must do structurally.
the oldest architecture in computing
The brain, treated literally as a computing architecture (comparable to Von Neumann) rather than metaphorically — one that predates and, per Vogels, still outperforms silicon on memory, energy efficiency, and learning-from-error.
trusted autonomy
The state in which an agent's demonstrated memory of a person's rules, preferences, and past corrections earns it license to act unattended — trust as a downstream product of good memory, not a separate design goal.
active forgetting
Deliberate, structural decay/compression/pruning of stored experience — a designed-in feature, not a failure mode — that keeps a system from spending more time sorting old context than doing useful work.

Propositions

explicitAI will automate some work, eliminate some job families, and consolidate some industries, following the pattern of every technology shift since the Industrial Revolution.conf: high
explicitStaying professionally relevant under AI depends on actively adapting and using new tools, not on the tools themselves.conf: high
explicitA biological brain is dramatically more energy-efficient than current silicon (GPU) computing at the same learning task.conf: high
explicitThe brain distributes knowledge across many local units (cortical columns), each building its own predictive model from direct experience and updating on prediction error.conf: high
explicitMemory consolidation happens in the background (in the brain, during sleep), distilling raw experience into structured long-term knowledge, strengthening what repeats and weakening what doesn't.conf: high
explicitThe brain is, literally, the oldest computing architecture — older than Von Neumann, the abacus, mathematics, and language.conf: high
explicitAn AI agent system engineered from practical constraints (Kiro Crew) converges on the same architectural pattern evolution produced in the brain: multiple distinct memory stores, background consolidation, and decay.conf: high
extendsThe Production Gap: 5 Patterns for Building Long-Running AI Agents · p8 — Turing Post argues long-term and working memory must be architecturally distinct stores; Vogels extends this with a concrete real-world example (Kiro Crew's markdown/vector/key-value split, mirroring the brain's distinct memory types) instantiating the same pattern.
explicitAny system that accumulates experience without active forgetting will eventually spend more time sorting old context than doing useful work.conf: high
refinesThe Production Gap: 5 Patterns for Building Long-Running AI Agents · p20 — Turing Post frames persistent memory as making behavioral drift 'a property of the system, not a bug'; Vogels reframes the same underlying dynamic positively — proposing active forgetting/decay as the designed countermeasure that keeps accumulated memory from degrading a system, rather than treating drift as an unavoidable cost.
implicitHuman-readable, human-editable memory (markdown files, an editable store) is an advantage AI agent memory has over biological memory, not merely a limitation.conf: medium
explicitPersistent, demonstrated memory of a person's rules and preferences is the mechanism by which an agent earns license to work autonomously and unattended.conf: high
implicitConvergent design (independently arriving at the brain's architecture from engineering constraints) is evidence that this architecture is close to optimal, not merely one option among many.conf: medium
implicitDeep, hands-on engagement with a new tool for weeks is necessary before an opinion about it (positive or skeptical) is credible.conf: medium

Arguments

The most effective response to AI-driven disruption is to personally adopt and deeply engage with AI tools, not to resist or avoid them. strength: medium

Explicit premises:

Every prior technology shift has automated some work while leaving room for those who adapt.
Vogels' own practice of extended, hands-on use of Kiro Crew is presented as the model of 'evolving.'

Implicit premises:

The reader has the time, access, and organizational latitude to spend weeks experimenting with new tools the way a senior technologist at a hyperscaler does.
Adaptation capacity, not current skill inventory, is what determines who remains relevant.

Inferences:

Organizations and individuals without slack time for exploratory tool use are structurally disadvantaged by this advice, even though the article frames it as universally available.

AI agent memory systems should be architected the way the brain is architected: multiple distinct memory stores, background consolidation, and built-in decay/pruning. strength: medium

Explicit premises:

The brain achieves structured memory, consolidation, decay, prediction and error-learning at a fraction of silicon's energy cost.
An independently engineered system (Kiro Crew), built from practical constraints alone, converged on markedly similar architecture (distinct memory types, background consolidation, decay).

Implicit premises:

Convergent evolution / convergent engineering toward the same solution is strong evidence that the solution is close to optimal, rather than one of several equally viable options.
What works for biological memory (which serves an organism's survival, embodied in a single continuous agent) transfers to what should work for software agent memory (which serves task completion across many stateless invocations, edited by an external user).
The Pong/GPU energy comparison, though about a wholly different task (control/motor learning) and computing substrate (spiking neurons on a chip vs. transformer training on GPUs), is treated as evidence about memory architecture, not just raw compute efficiency.

Inferences:

If active forgetting is architecturally necessary, agent memory systems that only ever add and never prune (a common naive implementation) will degrade in the same way an unpruned brain-like system would — a testable, falsifiable prediction the article doesn't itself pursue.

Agent memory is the mechanism through which agents earn 'trusted autonomy' to operate unattended. strength: low

Explicit premises:

Kiro Crew agents that learn a user's rules and preferences are described as earning trust to do work unattended.
Vogels reports a 'growing trust' with these agents he hasn't found with prior tools.

Implicit premises:

Persistent memory of preferences is sufficient (or at least primarily responsible) for earned trust, as opposed to track record on outcomes, verifiability of actions, or stakes of the task.
One person's (the architecture's own creator's) subjective growing trust generalizes to how trust should be extended by other users or organizations.

Inferences:

This argument implies that memory systems, not just capability improvements, are a lever for expanding the scope of agent autonomy an organization is willing to grant — a governance-relevant claim the article doesn't develop into any actual policy or safeguard.

Narratives

Vogels visits Cortical Labs (2023) and watches a network of lab-grown human neurons on a silicon chip learn to play Pong in a handful of attempts on 20 watts, versus a GPU's 700 watts and hundreds of thousands of iterations.
→ Grounds the abstract claim 'the brain is a better computing architecture' in a vivid, concrete, first-person-witnessed demonstration, making the energy-efficiency claim feel empirical rather than rhetorical.
A small engineering team builds Kiro Crew from practical constraints alone and arrives, apparently independently, at a memory architecture resembling neuroscience's account of the brain (distinct memory types, background consolidation, decay, pruning).
→ Implicitly argues that the brain-as-template thesis isn't just Vogels' analogy — it's borne out by convergent engineering, lending the analogy the authority of independent discovery rather than post-hoc pattern-matching.
One of Vogels' own Kiro Crew agents, while helping research the brain, proactively suggests building him a reading dashboard.
→ Illustrates 'trusted autonomy' and agent initiative in miniature — a small, low-stakes example standing in for the larger claim that memory-equipped agents can be trusted with unprompted, self-directed work.

Questions addressed

✅ Will AI take my job / make me obsolete?
Reframed: some work will be automated, but obsolescence is avoidable through active adaptation ('evolving') with the tools.
✅ What should AI agent memory architecture look like?
Modeled on the brain: distinct memory stores, background consolidation, graceful decay/pruning, and prediction-driven updating — evidenced by Kiro Crew's independently-arrived-at design.
⭕ Why is biological computation so much more energy-efficient than current silicon AI systems?
Named and illustrated (Cortical Labs' 20W vs. GPU's 700W) but never explained mechanistically — no discussion of spiking vs. dense computation, analog vs. digital substrates, or what (if anything) is portable from biological efficiency to silicon design.
⭕ How should trust in autonomous agents be calibrated and safeguarded as memory grows?
Not really addressed — trust is asserted as a natural consequence of memory ('agents that learn from you... earn the trust'), with only a passing personal caveat ('I still review what they do').

Gaps

incompleteThe article never explains the mechanism behind biological energy efficiency (sparse/event-driven spiking computation, analog signal processing, co-located memory and compute) versus why current silicon (dense matrix multiplication, separated memory and compute — the Von Neumann bottleneck the piece itself names) is so much less efficient. Without that mechanism, the reader can't tell what is actually transferable to future hardware or software versus what is a surface-level analogy.
Read next: Neuromorphic computing research (Loihi, TrueNorth, Cortical Labs' own published work on 'DishBrain'), analog in-memory computing, and critiques of end-to-end backpropagation as biologically implausible.
incomplete'Trusted autonomy' is asserted as a consequence of agent memory without any treatment of memory poisoning, drift, staleness, or manipulation — exactly the risks that would make memory an untrustworthy basis for autonomy rather than a trustworthy one. The article's own caveat ('I still review what they do') undercuts the trust claim without the piece noticing the tension.
Read next: Literature on agent memory drift and poisoning; governance frameworks for memory-driven autonomy (identity, audit trails, rollback of bad memories).
illogicalThe argument moves from 'a brain-like architecture was arrived at independently by an engineering team working from practical constraints' to 'the brain is the best architecture and always will be' — treating convergence as proof of optimality. But convergence could equally reflect that markdown files, vector databases and key-value stores are simply the standard toolkit available to any team solving this problem today, independent of any deep architectural truth about brains; the piece doesn't rule out this simpler explanation.
Read next: Philosophy-of-biology critiques of convergent-evolution-as-optimality arguments; software-engineering retrospectives on why certain storage patterns (hot/warm/cold tiers, vector + KV hybrid stores) recur across unrelated domains for purely practical reasons.
incompleteThe piece frames active forgetting/pruning entirely as a virtue (efficiency, avoiding context bloat) without acknowledging the corpus-documented failure mode of agent memory drift — that pruning and consolidation are themselves processes that can silently discard information a system later needs, or compress in ways that introduce error (the 'lossy consolidation' problem, analogous to false-memory formation in humans).
Read next: extract-turing-post-5-patterns-long-running-agents#p9 and #p20 on memory drift as a production failure mode when persistent state makes agent behavior a function of accumulated (and now also actively-edited/pruned) memory.

The oldest architecture in computing

Structure

Key terms

Propositions

Arguments

Narratives

Questions addressed

Gaps

Related documents