Claw learning loops must improve action capacity, not just retrieval

Type: kb/types/note.md

Simon Willison's account describes Claws as OpenClaw-like agents that generally run on personal hardware, communicate through messaging protocols, and act on direct instructions or scheduled tasks; the quoted Karpathy description also names orchestration, context, tool calls, and persistence. That description does not by itself entail a learning loop or retained knowledge about a user or project. This note concerns the subset whose persistence does retain domain, user, or project context for later tasks. For such a Claw, the system-level learning objective is contextual competence: the capacity to act appropriately given what it has retained. Retrieval remains an essential subsystem objective, but it is not a sufficient success criterion for the Claw as a whole.

This distinction separates action capacity from action outcome. Retained knowledge can expand the range of actions a system is equipped to perform without guaranteeing that any particular action will succeed. Execution, classification, communication, planning, and precedent recognition all depend on accumulated context, but their outcomes also depend on the model, prompt, tools, permissions, and environment.

Action capacity needs action-guiding artifacts

Reference facts and relationships are not the only retained material that can improve future work. A Claw may also need preferences about how work should be done, procedures learned from successful executions, and precedents that preserve earlier judgments and their rationale. Voice guides and domain models are further examples of action-guiding artifacts, not evidence that the document type system itself must change.

Koylanai's Personal Brain OS provides one anecdotal example. Its self-reported design separates voice and brand files, goals and values, process-encoding skills, and append-only experience, decision, and failure logs across isolated modules. Treating these as examples of preferences, procedures, precedents, and voice material is this note's abstraction, not the source's taxonomy. The report establishes that one such design exists, not that practitioner-built Claws converge on it.

Existing mutations can learn different content

The KB learning loop proposes extracting, synthesising, reformulating, regrouping, and retiring artifacts. Those operations can already express preference codification, procedure capture, and precedent consolidation. What changes is not necessarily the mutation vocabulary, but which candidates are proposed and what evidence licenses their promotion.

For example, repeated user choices may suggest a preference, while successful traces may suggest a procedure. Neither should become authoritative merely because it appeared once or correlated with success: choosing what to learn requires both validity and learning-value gates.

Evaluation needs contribution attribution

At the Claw level, the relevant question is whether behavior became more contextually competent. At the KB-subsystem level, the questions are narrower: Did the needed artifact exist? Did it represent the lesson faithfully, reach the relevant context, and causally affect the action? These questions preserve a modular diagnostic boundary. An action failure licenses a KB mutation only when the failure can be attributed to missing, misleading, unavailable, or behaviorally inert retained knowledge rather than to the model, prompt, tools, permissions, or environment.

Existing systems illustrate both sides of this boundary. Agent Workflow Memory extracts workflows from successful traces for later tasks, while Voyager promotes successful programs into a reusable skill library. In the behavioral-faithfulness study, causal perturbations changed behavior more consistently for raw trajectories than for condensed experience, and weak dependence on condensed content persisted when it was the only experience supplied. The result does not prove that the model failed to read the text or that condensed experience never helped. It does show why presence and retrieval alone do not establish semantic uptake or learning impact.

Retrieval learning is therefore a valid layer, not a rival objective. It can measure artifact availability, fidelity, and selection. The broader action outcome indicates whether the Claw's competence improved; attribution tells its learning loop which subsystem, if any, should change.

Open Questions

  • What practical intervention or counterfactual can establish that a retained artifact caused an action improvement rather than merely co-occurred with it?
  • Which end-to-end failures can be attributed cheaply enough to license a KB mutation rather than a model, prompt, tool, permission, or environment change?

Relevant Notes: