Case packet
Neutral case identifier: case-41428517ec8382
The possible directed relationship from Artifact A to Artifact B is under review.
Artifact A
Agent memory needs discoverable, composable, trusted knowledge under bounded context
Agent memory is not valuable because something was stored after a session. A durable fact, claim, preference, procedure, rationale, example, or learned constraint counts as useful memory only when it can change future agent work. The success criterion is [contextual competence]: the agent's ability to answer, classify, plan, communicate, edit, and choose behavior appropriately under bounded context.
That makes memory broader than retrieval but narrower than "anything persistent." A memory can be retrieved and still fail if it cannot be used. It can be stored and still fail if the agent never encounters it. It can be true and still fail if the agent cannot tell when to rely on it. Remembered knowledge earns its place by improving future action capacity, not by existing in a store or matching a query.
This note is about remembered knowledge, not the whole memory system. Capture, activation, provenance, authority, lifecycle, evaluation, and governance are necessary machinery around memory. The first-order quality test for the remembered material itself is simpler: it must at least be discoverable, composable, and trusted. These are a minimal basis for artifact quality, not an exhaustive ontology of every property a memory system must maintain.
Three properties of useful agent memory
Discoverable means the agent or context engine can find the remembered material when it matters without scanning everything. The failure mode is inert memory: the system has a relevant lesson, preference, rule, or rationale, but it never enters the working context where it could change the outcome. Discoverability can come from titles, descriptions, tags, indexes, retrieval keys, activation cues, source metadata, or placement in an always-loaded control plane. The common requirement is not a particular retrieval method but a usable handle from future situations to the remembered material.
Composable means the remembered material can combine with task context and other memories to support new reasoning or action, or can be applied as a coherent unit where direct reuse is the point. The failure mode is isolated memory: a fragment is retrievable, but the agent cannot tell its scope, relation, priority, consequence, or application conditions, so it cannot use the fragment as a premise or apply it as a unit. Composable memory has enough shape to answer "what does this apply to?", "what does it depend on?", and "what should change if I believe or use it?" A stored preference, a source-derived claim, a reusable template, a test, a skill, and an architectural decision all need different shapes, but each must be usable with other remembered material rather than trapped as a disconnected fact.
Trusted means the agent can rely on the remembered material at the right level of confidence without redoing the original work every time. The failure mode is memory as noise: the agent either ignores the memory because its authority is unclear, or over-applies it because its provenance, currency, and scope are missing. Trust is not the same as truth. It is calibrated reliance: enough rationale, source connection, status, validation, ownership, or review signal for the memory's role. A tentative observation can be trusted as tentative; a validated rule can be trusted as a rule.
These properties are interdependent. Composability depends on discoverability because unavailable memory cannot participate in reasoning. Composability also depends on trust because unreliable premises poison downstream conclusions. Trust depends on discoverability because memory that cannot be found cannot be challenged, corrected, reviewed, or retired. Discoverability is therefore the entry condition, but it is not sufficient: memory that is findable but unusable or untrusted still does not improve contextual competence.
Other memory requirements operate around this basis
Discoverability, composability, and trust do not exhaust the requirements for a working memory system. They name the minimum artifact-quality basis: if remembered material lacks one of them, it cannot reliably improve future action. Other requirements become peer properties only if remembered material can satisfy the triad yet still fail as remembered material. Otherwise, they operate around the basis by preserving it, operationalizing it, or deciding when it should matter.
Context economy is one example. Useful memory must be economical enough to load or compile into the future situation where it matters. A memory can be nominally discoverable but too noisy to select, nominally composable but too bulky to combine, or nominally trusted but too expensive to verify or carry. Activation, lifecycle, governance, evaluation, provenance, authority, and capture play similar surrounding roles: they are not additional peer properties of the remembered material, but system-level mechanisms that keep the artifact-quality basis useful under bounded context.
Relevant Notes:
Artifact B
Claw learning loops must improve action capacity, not just retrieval
A [Claw] accumulates context and acts on a user's behalf. Its system-level learning objective is therefore contextual competence: the capacity to act appropriately given what it has retained about the domain, user, and project. 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 stores preferences, procedures, precedents, and voice material separately. This establishes that such a 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. Yet retained experience can be present without affecting behavior. Storage and retrieval alone therefore do not establish learning impact, as the [behavioral-faithfulness evidence] shows.
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:
Under-review context phrase
contextual competence is the success criterion for learned material, not retrieval accuracy