Self-improving systems

Type: kb/types/tag-readme.md

A self-improving system makes operative changes to its own behavior-determining organization in response to evidence about an improvement objective. The definition is deliberately broad — a weight-level learner qualifies, and so does a dev team. The distinction that carries the design information is whether the improvement pathway is reflective — routed through a self-representation — or not. Most systems in agent-memory-systems mine their traces for lessons and load them into later runs — a bid at the reflective kind that succeeds only when the loaded lessons function as a representation of the system's own behavior, not just accumulated knowledge.

Why reflection might pay

The case for reflective pathways is a four-step chain; only the first step is architectural rather than conjectural.

1. Reflection makes retained lessons addressable: later rounds can read, criticize, and selectively revise them — reflection buys addressability.
2. An addressable lesson is second-order: it can reject a prior commitment outright, where a gradient can only nudge it — reflection makes retained lessons second-order.
3. The payoff conjecture: reflection may improve sample efficiency under structured shifts. This is Commonplace's own bet, not a result; the note carries the prediction and the test design.
4. The standing discount: retrieval is best-effort, and a lesson that never surfaces contributes nothing — retrieval failure is reflection failure.

The three gradings

Since membership is cheap, a system is characterized by where it sits on three independent gradings:

• Retention form — operative, cumulative, or addressable. The chain above is this axis's payoff; the machinery is the reflective system's causally connected self-map.
• Coverage — reflective coverage is graded across representational forms.
• Closure — a methodology governs its own extension only as far as it settles the meta-decisions it raises.

Autonomy — how much of the pathway runs without a person — is a fourth thing worth reporting for any system, but it is not tracked as a fourth placement here: under Commonplace's strictly computational reflective-system boundary, a pathway that is reflective at all is thereby also autonomous, since admitting a human into the boundary would only trade that precision for a separate axis. Non-reflective pathways are reported human-inclusive or autonomous directly, using the base definition's own boundary-relative grading. What costs is warrant, trusting what an unattended loop accepts: warranted autonomy is bounded by oracle domain.

Example placements

• Ashby's Homeostat — the floor: non-reflective, fully autonomous, nothing accumulates.
• Parametric self-improvers (self-play policies, agents fine-tuned on their own trajectories) — compounding without addressability; the dominant paradigm and the conjecture's comparison baseline.
• The Gödel machine — reflective, autonomous, proof-gated: the strongest oracle, paid for in domain.
• Commonplace itself — pathway-mixed: reflective and autonomous where a validator or agent consults an explicit self-representation, non-reflective and human-inclusive where a maintainer notices what's worth fixing or judges the shape of a fix.

Read every claim at its stated strength: the definitions are stipulated and revisable, the reflective advantage is a hypothesis, and whether it holds is the open empirical question the chain above sharpens.

Further notes under the tag

• Measuring autonomy well enough to see it improve is an open problem — the per-function grading above tells you where a system sits, but not whether it is getting more autonomous or how it compares to a differently-decomposed system.
• Behavior-determining organization, operative change, evidence bearing on an improvement objective — the definition's three base terms.
• The definition classifies its boundary cases without ad hoc exceptions — nine inclusion tests, from gradient learning to accidental self-modification.
• A proposal-selection loop requires search, evaluation, and operative retention and false-positive acceptance becomes operative — the generate-evaluate-retain subtype and why evaluation is its terminal filter.
• Improving an agentic system crosses the prose-symbolic boundary — reliability gains move behavior between prose and code, so coverage of one form cannot carry them. The agent-memory-system reviews and the comparison matrix run on this vocabulary.
• Reach assessment — the semantic judgment reflectivity's structural requirements do not supply; the sample-efficiency conjecture and second-order rescoping both depend on it, and current evaluators seem to have it with no theory for why.
• Formal symbolic systems assess reach only through causal and proof obligations — reach assessment isn't LLM-exclusive: causal inference and proof search give traditional symbolic systems a formal route to it once a claim's generality is encoded as a checkable obligation.

Related Tags

• foundations — the broader core theory this sits inside
• constraining — closure is a constraining property of methodology-as-input
• computational-model — reflection and intercession as computational concepts generalized to socio-technical boundaries