Explicit retention provides direct targets for selective revision

Type: kb/types/note.md · Tags: learning-theory, self-improving-systems, agent-memory

A retained theory, rule, or procedure can make a commitment an object the learner retrieves, criticizes, and revises. This is the practical advantage of explicit retention: it supplies a direct target for a named change. It does not establish that explicit artifacts are the only way to learn, transfer a judgment, or govern behaviour.

Three properties must be assessed separately. Durability means state survives across the horizon of the learning claim. Writability means the system can change it through its permitted operations. Addressability means those operations can target the relevant commitment, rather than only replace or probe the component as a whole. Reflection buys addressability, but neither a readable file nor a numerical parameter guarantees a clean boundary around a semantic commitment.

Compare operations, not opaque and readable substrates alone

Retention form Durability and writability Handle on a commitment
Context-conditioned state Depends on runtime persistence; discarded state does not survive the episode The retained transcript may be editable even when the state it induces is not directly inspectable
Pinned model parameters Durable over the run, not writable under the pin Prompting or probing changes the conditions of use, not the parameters
Parameters with an allowed update path Updates can persist A targeted edit needs evidence that it changes the intended judgment without unacceptable collateral effects
Human expertise inside the boundary Can persist and change through practice Articulation and behavioural tests expose some commitments; the person's whole competence is not a versioned artifact
Natural-language and symbolic artifacts Durable when retained and writable when permissions allow Text spans, rules, functions, and fields provide direct edit targets; their consequences still need checking

These are operation profiles, not permanent rankings of representational forms. A read-only note is not writable. A vague theory paragraph may not isolate one assumption. A retained context can affect later work even without a separately written theory. A numerical representation with a tested commitment-level editing interface should be assessed by that interface, not excluded by its form.

The distinction between a transcript and the competence it induces is useful. Editing a sentence is direct control over the text, not guaranteed selective control over the resulting judgment. The same limitation applies to a retained theory: its apparent locality must be checked against the behaviour of the system that consumes it. Reflective coverage therefore reports the operations available on each part.

What follows under the fixed-model premise

When model parameters are pinned and the declared writable surfaces are natural-language and symbolic artifacts, retained changes must use those surfaces. This follows from the declared update rules, not from a theorem that other substrates cannot learn. A cached state or external record also needs its own declared persistence and consumption path.

Commonplace's declared frame places provider weights outside the revision boundary. Changing a model binding there edits a configuration request; it does not edit the provider's weights. That describes one system's available operations, not a general limit on learning systems.

Explicitness supports one route to transfer

A human can externalize a criterion as a rule, rationale, example, or test that computation later consumes. A named criterion makes it easier to compare what was transferred and to revise the record when its use fails. This is useful for methodological closure, which asks what a retained method settles.

Computational transfer need not follow that route. A model may already supply the required judgment, infer it from examples, or acquire it through an allowed parameter-learning process. Such a transfer may be harder to inspect commitment by commitment. It can still be governed through outcome checks, limited authority, regression tests, and rollback. Lack of a directly editable criterion is not lack of all governance, and automatic execution is not proof that a criterion has become explicit.

What would establish the advantage

At a selected decision, identify the commitment to change and compare the available editing routes. Measure whether the intended later behaviour changes, what collateral behaviour changes, whether the update can be reversed, and the cost of diagnosis and validation. Explicit retention earns its place when its direct targets make that process more useful at acceptable cost.

An equally selective numerical or reconstructed representation would defeat an exclusivity claim. A readable edit that repeatedly changes unrelated judgments would defeat the assumed locality of that artifact. Neither result would deny that retained state can support learning.

Scope

Addressability concerns the commitment and the operations the claim actually needs. It is not complete transparency of the learner. Retained episodes, examples, theories, and programs can supply different information and different edit targets; retaining an episode can preserve details that a distilled rule omits. None guarantees that later interpretation recovers those details correctly.

Relevant Notes: