1. From bot to agent: the jump

An LLM without actuators is a bot: it speaks well but does not act. When you add sandboxed tools it becomes an agent: it speaks and acts, but only with the actions the designer hard-wired for it. The next step — and the goal of this document — is to make it capable of building itself new actuators when the existing ones aren't enough, within safety limits.

The metaphor: an LLM with fixed tools is like a brain connected to hands that can only make certain gestures. If the task requires turning a key and the hands can only press buttons, the agent fails. A good agent, at that point, fabricates the capability to turn the key — in our case by writing a small code applet — but does so under supervision.

We call these applets neurons because they have three properties that echo across the neuroscience literature:

1. They persist: once approved, they survive the session that generated them. They come back at the next need.
2. They activate: facing an input that concerns them, they "fire" and produce an action.
3. They form networks: they invoke each other; the frequent paired use of two neurons links them (a synapse strengthens).

2. What a neuron is

A neuron is the materialisation of a capability Mykleos did not have, synthesised on demand, approved by the user, and then reusable.

3. The synthesis cycle of a neuron

A neuron does not arise spontaneously. It is born from an explicit failure: the agent has tried all known tools, none produced the required result, and the user has not granted permission to abandon the task.

4. The Darwinian law: selection by gap reduction

A neuron is not a vested right. The initial approval gives it birth, but it stays in the library only if it produces utility. Utility has a precise definition: reduction of the gap between Mykleos's current capability and the capability needed to reach the goal. The rest follows: those that close the gap live, those that don't die, the environment — that is, the goals Mykleos pursues — selects.

Definition of Gap, utility, fitness

Corollaries

• Used ≠ useful. A neuron may be invoked because "it seems right" and fail to close the gap. What matters is the measured effect, not apparent plausibility.
• Competition. Two neurons with comparable fitness compete for the same goal: the cheaper one wins (CPU time, LLM cost, I/O calls).
• Fossils. A neuron with persistently negative fitness enters quarantine after N strikes, then extinction. The file stays for forensics, the signature is revoked.

Two modes for goal origin

The goal that measures fitness can arise from outside (the user asked for something) or from inside (Mykleos notices a recurring gap and wants to close it). They are qualitatively different and need different gates.

How Gap is measured, in practice

The selection arena

The Darwinian cycle, phase by phase

Caveats from the literature

• The LLM's self-evaluation "did I reduce the gap?" is optimistic. Where possible, use objective metrics (time, exit code, expected match), not "an LLM judging an LLM". See Huang et al. (2023), "Large Language Models Cannot Self-Correct Reasoning Yet".
• Selection vs exploration. A dormant neuron could be useful for rare goals. The retriever cannot prune all that lacks fitness: it uses a bandit (UCB-like) that reserves a quota of exploratory invocations.
• Memory poisoning. If an attacker artificially inflates the fitness of a malicious neuron, it dominates the retriever. Defence: every fitness event is signed by the caller, verifiable in the audit log, and aggregated only after a confirmation window.
• Experience vs dogmatism. Fitness is summed experience, not certainty. The decision "invoke this neuron" still has to go through Policy: a neuron with high fitness but which touches a forbidden path stays blocked. The Laws win over the numbers.

5. Synapses: how neurons talk to each other

A single neuron is one more tool. A set of neurons that call each other is a network. The structure of that network — who calls whom, how often, with what success — is the equivalent of a behavioural memory of the system.

How synapses form

1. Declared: a neuron's manifest explicitly lists which other neurons it might call. Maximum calling constraint.
2. Observed: every co-activation (neuron A calls B in the same trace) increments a counter. The real synapses emerge from use, not from declaration.
3. Decay: counters that don't grow for 30 days scale down linearly. Pairs of neurons actually used together stay strong, others fade away.

What they CANNOT do, among themselves

• A neuron cannot call another neuron not declared in its manifest. The registry filters.
• A neuron cannot create other neurons. Only the synthesizer of the main agent can. (Anti-runaway lock.)
• A neuron cannot escalate its capabilities at runtime. They are fixed at signature.

6. The three-tier memory

We need three time horizons, with three different scopes. The distinction is not decorative: each tier has a prompt-injection criterion, a persistence criterion, a compression criterion.

Immediate (scratchpad)

Lives in RAM, in the agent process alone, for the duration of the current request. It contains: intermediate thoughts ("I need to read the log first"), partial tool outputs, the plan yet to execute. When the action ends, it is discarded or — if it contains something worth keeping — promoted to medium memory via distillation.

Medium (functional)

An SQLite table in workspace/memory/medium.db. It contains the history of the current conversation and recent coherent tasks. It has automatic retention: after N days or M messages, compression (via LLM summary) and archival. Parts that during compression turn out to be stable and general are proposed for promotion to long.

Long (constitutional)

The markdown files of the workspace: SOUL.md (the Constitution — the four Laws), IDENTITY.md, USER.md, the stable part of MEMORY.md. This block is always at the head of the system prompt, with Anthropic/OpenAI prompt caching so it isn't paid for on every call. It is also the only tier Roberto edits directly with a text editor.

7. The Four Laws of Mykleos

Long memory has a core not editable even by Roberto without an explicit rite: the Constitution. These are Asimov's laws reworked for a household butler. They live in SOUL.md and are the first content every LLM call sees.

The Laws are numbered and ordered: on conflict, the lower-numbered Law wins. They are injected as a cacheable block into every prompt, before the rest of long memory.

8. Quotas, kill-switch, quarantine

9. Where neurons live: impact on the 4 layers

With respect to the starting architecture (intro §3), neurons do not create a new layer: they inhabit layer 4 (Workspace & Tool) and do so with the same constraints as native tools.

10. Anatomy of a neuron on the filesystem

workspace/neurons/
└── nas-ping/
    ├── manifest.yaml       # purpose, capabilities, quotas, declared synapses
    ├── neuron.py           # the body (Tool Protocol)
    ├── test_neuron.py      # the birth test
    ├── signature           # HMAC over body+manifest, issued by myclaw root
    └── journal.jsonl       # usage counters, latest outcomes, errors

Minimal example of manifest.yaml:

name: nas-ping
version: 1.0
born: 2026-04-21T22:14:00Z
approved_by: roberto
scope: "check home NAS reachability via ICMP ping + arp"
capabilities:
  - net.ping: { hosts: ["192.168.1.50"] }
  - shell.exec: { allowlist: ["/usr/bin/ping", "/usr/sbin/arp"], timeout_s: 5 }
quotas:
  cpu_s: 2
  ram_mb: 64
  io_writes: 0        # read-only neuron
synapses_declared:
  - tg-send           # may notify via Telegram
  - log-parser        # may read its output
signature_algo: hmac-sha256

11. What a neuron is NOT

• Not a plugin. No installer, no marketplace. Born from the agent's need, here at home, for Roberto's use case.
• Not a sub-agent. No reasoning loop of its own, no own memory. It is a tool, with extra metadata.
• Not a microservice. Doesn't run as a long-lived process, has no ports, exposes no API to anyone except the main agent.
• Not a substitute for careful programming. If the required capability is foreseeable and stable, better implement it as a native tool. Neurons are for gaps emerged from use, not for planned features.
• Not a learning mechanism in the ML sense. It does not update weights nor fine-tune. It evolves structurally (new nodes, new synapses), not parametrically.

12. Updated roadmap

Neurons enter the roadmap after phase 4. Before knowing how to generate them, Mykleos must know how to execute tools safely.

Microdesign documents this extension adds

• neuron.html — structure, manifest schema, signature
• synthesizer.html — synthesis pipeline, gates, rejection
• synapse.html — graph, counters, decay
• constitution.html — the 4 Laws, the modification rite
• memory.html (already planned) — now with 3 explicit tiers

I am adding them right away to the microdesign index (in Italian).

Keep reading

Mykleos — Neurons, Synapses and Memory v1.1 — 2026-04-21
Extends Architecture — Introduction v1.