Extractable theory notes from agent memory design

Target: Designing a Memory System for LLM-Based Agents

Purpose: identify claims embedded in the synthesis note that could become standalone theoretical notes in kb/notes/, using the notes collection's quality bar: claim title, reach beyond the design study, composability as an imported premise, clear boundary conditions, and no dependence on the whole four-layer architecture.

Summary

The source note is doing two jobs at once:

  1. It proposes an architecture for agent memory: trace, observation, episode, library; role-split retrieval; typed cue indexes; extraction taxonomy.
  2. It contains several general claims about memory, activation, promotion, source-of-truth control, and trace-derived learning.

The architecture can remain a synthesis note. The extractable material is the set of claims that another note could cite without importing the full architecture. The strongest candidates are not "parts of the architecture"; they are mechanisms that apply across memory systems, KB maintenance, review systems, and agent runtime design.

The stronger grounding found in the grounding report is that these candidates sit at the intersection of three theory clusters:

  • Context engineering: memory is useful only when a context engine selects, frames, and injects it into a bounded call. This grounds claims about capture posture, cue activation, progressive disclosure, and trace loading.
  • Readable-artifact learning: session traces become durable capacity through distillation, constraining, and codification. This grounds claims about promotion, compiled cues, negative results, and procedure extraction.
  • Oracle theory: extraction and promotion differ by signal strength. Corrections have strong interactive oracles; discoveries and broad syntheses have weak or delayed oracles. This grounds claims about build order, behavioral faithfulness, and promotion authority.

This suggests a useful filter: extract theory notes when the claim names a reusable mechanism in one of those clusters. Keep material inside the synthesis note when it only justifies the proposed four-layer architecture.

Best extraction candidates

Store-everything is a capture posture, not a memory design

Likely title:

Store-everything is a capture posture, not a memory design

Claim:

Retaining raw traces broadly preserves optionality for later redistillation, but it does not determine what the memory system should extract, promote, retain, redact, or evaluate. Those decisions must be governed by the output the memory system is meant to improve.

Stronger grounding:

This should be grounded first in context engineering, not storage economics. Context efficiency is the central design concern in agent systems explains why storage and context have different cost models: the hard problem is bounded context assembly, not persistence. Agent runtimes decompose into scheduler context engine and execution substrate locates broad trace retention on the execution substrate, while extraction/loading belong to the context engine and learning loop. Agent memory is a crosscutting concern, not a separable niche gives the most direct frame: "memory" decomposes into storage, activation, and learning.

Why it deserves a separate note:

The source note starts from the store-everything premise, then softens it by calling it a bet. The previous critique correctly notices that this is still too input-driven. A standalone note would reconcile the source note with the output-driven memory workshop: broad capture is a substrate choice; memory design is a target-relative policy.

Reach:

This applies beyond agent memory: logs, research notes, observability traces, full transcript retention, source snapshots, and workshop artifacts all benefit from broad capture but still need output-governed extraction.

Boundary:

The claim should not argue against broad trace retention. It should argue against confusing retention with value. "Store everything eligible under retention/redaction policy" is compatible with output-driven extraction.

Existing overlaps:

Suggested links:

System-definition cues need behavioral faithfulness tests

Likely title:

System-definition cues need behavioral faithfulness tests

Claim:

A memory cue that fires and enters context has not closed the activation gap unless it causally changes the agent's later action in the intended direction. Cue quality must be evaluated by behavioral uptake, not by retrieval success alone.

Stronger grounding:

The cue is a system-definition artifact, so this belongs in the behavior-change half of continual learning. Continual learning's open problem is behaviour, not knowledge grounds why a retrieved fact and a fired policy cue have different success criteria. The readable-artifact loop is the tractable unit for continual learning explains why prompt snippets, rules, tests, and tools are the practical substrate for this kind of behavior change. Oracle strength spectrum explains why the evaluation should move from "cue was present" to an oracle that discriminates changed behavior from unchanged behavior.

Why it deserves a separate note:

The source note has a strong activation design: cue match, priority arbitration, commitment. But "commitment" currently depends on imperative framing, checkpoint insertion, and contradiction surfacing. Recent evidence in the KB says live context can still be ignored. A standalone note would upgrade the activation theory: the unit of evaluation is not "was the cue retrieved?" but "did the cue change behavior for the intended reason?"

Reach:

This applies to always-loaded instructions, review gates, safety policies, prompt snippets, skills, guardrails, and generated memories. Any system-definition artifact that claims to steer action needs a behavioral test.

Boundary:

This is strongest for high-priority or always-loaded behavior-changing cues. Low-risk advisory context may not justify expensive ablations.

Existing overlaps:

Suggested links:

Library-derived cues should be compiled views

Likely title:

Library-derived cues should be compiled views

Claim:

When a cue index entry is generated from a library artifact, the library artifact should remain the source of truth and the cue should carry provenance, source hash, and regeneration rules. Otherwise the cue becomes a divergent behavior-changing policy surface.

Stronger grounding:

The important addition from the grounding pass is role: a library-derived cue is not just a derived summary, it is a derived system-definition surface. That makes drift more dangerous than ordinary documentation drift because the cue can change behavior even when the explanatory source has changed. Axes of artifact analysis grounds this role distinction; System-definition artifacts are crystallized reasoning under context scarcity explains why those artifacts exist; Always-loaded context mechanisms in agent harnesses shows the broader family of behavior-shaping context surfaces that can drift from their sources.

Why it deserves a separate note:

The source note's "library to observation backflow" is important but under-specified. It says a plain library artifact often needs a companion cue so it can fire in the right situation. That creates a source-of-truth problem: library prose and cue index can drift. This claim is general and operationally useful outside agent memory.

Reach:

Applies to compiled indexes, generated review gates, AGENTS.md excerpts derived from notes, quickstart files derived from canonical guides, generated lint rules derived from policy, and any fast path derived from a deeper explanation.

Boundary:

Directly authored candidate cues are different. The compiled-view rule applies once the cue claims authority from a library artifact.

Existing overlaps:

Suggested links:

Episodes are retrospective memory; workshops are active memory

Likely title:

Episodes are retrospective memory; workshops are active memory

Claim:

Retrospective episodes and active workshop artifacts solve different memory problems. Episodes compress what happened after a bounded effort; workshops hold work in motion with state, dependencies, expiration, and unresolved alternatives.

Stronger grounding:

The grounding pass clarifies that this is a runtime-boundary claim, not just a document-type distinction. Episodes are distilled retrospective artifacts for later retrieval; workshops are active state on the work surface. In the runtime decomposition, both live on the execution substrate, but they feed the context engine differently: episodes are loaded as memory; workshop artifacts are read and mutated as current work state. This prevents the four-layer architecture from accidentally absorbing the active work surface into retrospective memory.

Why it deserves a separate note:

The source note's episode layer is retrospective. The KB already has a workshop-layer theory, but the relation between "episode" and "workshop" is not yet explicit. The source note risks treating active work as trace or episode, which loses stateful process structure.

Reach:

This applies to KB design, agent runtime memory, project management, decision-record production, review runs, research workflows, and any setting where work has both live state and later narrative recall.

Boundary:

The note should not propose adding a fifth memory layer by default. It should clarify that the four-layer design describes retrospective memory; active workshops are adjacent work-surface artifacts that can later produce episodes and library notes.

Existing overlaps:

Suggested links:

Negative results are memory artifacts because final artifacts hide discarded paths

Likely title:

Negative results are memory artifacts because final artifacts hide discarded paths

Claim:

Durable work products usually preserve the chosen path, not the rejected alternatives. A memory system needs first-class negative-result records because the value of "why not X?" is highest when X becomes tempting again.

Stronger grounding:

The grounding pass adds two useful frames. First, this is a knowledge-role memory artifact with a possible system-definition companion: the negative-result record answers "why not?", while a cue can fire when the rejected path is proposed again. Second, promotion should follow oracle strength and reach: the failed path needs enough evidence, recurrence likelihood, or consequence to justify a durable artifact.

Why it deserves a separate note:

The source note treats negative-result preservation as a use case. The claim has enough reach to stand alone: final artifacts systematically omit discarded paths, so a separate memory artifact is needed to prevent rediscovery loops.

Reach:

Applies to research, design, architecture, policy, product decisions, debugging, source selection, and agent behavior. It is not specific to the four-layer architecture.

Boundary:

Not every failed attempt deserves promotion. The note needs a promotion standard: the rejected path should be likely to recur, costly to retry, or useful as evidence for a broader claim.

Existing overlaps:

Suggested links:

Good candidates that may be revisions instead of new notes

Atomic memory and narrative memory need different retrieval interfaces

Possible title:

Atomic memory and narrative memory need different retrieval interfaces

Claim:

Atomic observations answer "have we seen this before?"; episodes answer "what happened when we tried this?" The former wants lookup and trigger matching; the latter wants narrative compression and contextual reading.

Reason to hesitate:

This is currently the strongest internal argument for the four-layer architecture. It might belong inside the source note unless another note needs to cite the distinction independently. If extracted, it should weaken the layer-count claim: the distinction requires different interfaces, not necessarily different physical storage layers.

Added grounding:

The stronger extraction may be broader than this title: memory layers are progressive distillation interfaces. Pointer design tradeoffs in progressive disclosure grounds the general mechanism: descriptions, observation summaries, episode goals, and full artifacts are different-resolution pointers used to decide what to load. Distillation grounds the transformation between resolutions. This makes the note less about "atomic versus narrative" and more about matching artifact shape to consumer task and context budget.

Quality risk:

The claim is solid but may be too close to architecture justification unless generalized beyond the four-layer design. If extracted, avoid arguing for a fixed trace/observation/episode/library stack. Argue instead that memory systems need distinct retrieval interfaces when their consumers ask different questions: "has this happened?", "what happened?", "why did we choose?", and "what should I do now?"

Suggested links:

Typed cue indexes are on-situation context injection

Possible title:

Typed cue indexes are on-situation context injection

Claim:

Typed cue indexes are the action-triggered counterpart to on-reference loading and on-invoke skill loading. They inject system-definition context when a situation or proposed action matches a stored trigger, rather than waiting for a direct query.

Reason to hesitate:

The exact "typed cue index" mechanism may be too implementation-shaped for a theoretical note. But the broader loading category, on-situation injection, is reusable. It connects the source note's cue machinery to the KB's context-engine theory.

Added grounding:

Agent statelessness means the context engine should inject context automatically already proposes "on reference" injection. Always-loaded context mechanisms in agent harnesses distinguishes always-loaded, on-demand, and capability-description surfaces. This proposed note would add a missing row to the loading hierarchy: always, on reference, on situation, on invoke, on demand.

Quality risk:

The note should not overclaim that typed cues solve activation. It should stay at the context-engine mechanism level and defer success criteria to System-definition cues need behavioral faithfulness tests.

Suggested links:

Extraction taxonomy should follow oracle strength

Possible title:

Extraction taxonomy should follow oracle strength

Claim:

Session-log extraction types should be ordered and operationalized by the strength of the signal that justifies extraction and promotion. Corrections, silent failures, preferences, procedures, discoveries, decision provenance, and negative results differ less by topic than by oracle quality, recurrence needs, and promotion risk.

Reason to hesitate:

This may be a revision to the source note's extraction taxonomy rather than a standalone note. Extract it if the oracle-strength ordering becomes useful outside agent memory, for example in review-system learning, workshop-to-library promotion, or KB maintenance automation.

Added grounding:

Oracle strength spectrum provides the primary vocabulary. Trace-derived learning techniques in related systems supplies evidence that trace-mining systems converge on similar stages but stall at evaluation, trust, persistence, and retirement. Memory management policy is learnable but oracle-dependent explains why learned extraction/promotion policies only close cleanly when the task supplies a strong oracle.

Quality risk:

The note could become a catalogue. It needs a sharp claim: extraction schemas and build order should be derived from oracle strength, not from an intuitive list of memory types.

Suggested links:

Graduation should wait until retrieval cost exceeds maintenance cost

Possible title:

Graduation should wait until retrieval cost exceeds maintenance cost

Claim:

Promoting an observation into a durable artifact creates maintenance obligations. Late graduation keeps storage cheap; premature graduation creates stale policy and curation debt.

Reason to hesitate:

This is valuable, but it may overlap with existing maintenance and entropy notes:

It may be better as a revision to one of those notes or as a short companion only if the "retrieval cost versus maintenance cost" formulation becomes reusable.

Added grounding:

The grounding pass adds a role-sensitive version of the claim. Knowledge-role graduation creates curation and retrieval obligations; system-definition graduation also creates behavior-change risk. That means the threshold should be higher for always-loaded instructions, guardrails, compiled cues, and tests than for ordinary knowledge records. Axes of artifact analysis, Methodology enforcement is constraining, and Oracle strength spectrum are the useful grounding links if this becomes standalone.

Reach reveals itself through accumulation

Possible title:

Reach reveals itself through accumulation

Claim:

An observation's generality is often unknowable when first captured. Multiple low-reach observations can later reveal a high-reach pattern, so capture and promotion should be separated.

Reason to hesitate:

This is elegant and probably true, but it partly restates the source note's store-broadly/promote-selectively position. It could become the theoretical core of "store-everything is a capture posture," or a separate note if the reach mechanism becomes central elsewhere.

Suggested direction:

Do not extract first. Fold it into the store-everything/capture-posture note as one mechanism.

Added grounding:

This is where first-principles reasoning selects for explanatory reach over adaptive fit should enter if the claim is extracted later. The memory-design version is: low-reach observations are often only distinguishable from high-reach patterns after accumulation. Quality signals for KB evaluation can provide possible weak signals for when reach has become visible through links, recurrence, and neighborhood coherence.

Silent failures should be extracted before final success normalizes them

Possible title:

Silent failures should be extracted before final success normalizes them

Claim:

When a task succeeds through fallback, retry, or workaround, the final answer hides degraded guarantees and tool defects. Session-log extraction should treat these as first-class repair or policy candidates.

Reason to hesitate:

There are already strong observability notes:

This may be better as a revision to the first note or a prescriptive extraction rule in instructions, unless we want a memory-specific note about "silent-failure extraction."

Added grounding:

The oracle-strength grounding makes this a stronger candidate than before. Silent failures have visible operational signals, errors, retries, fallback paths, degraded guarantees, but their significance still requires judgment. That puts them between hard-oracle corrections and weak-oracle discoveries. Spec mining is codification's operational mechanism also explains the graduation path: repeated silent failures can become checks, recovery rules, or deterministic guards.

Suggested links if extracted:

Lower-priority or keep-inside candidates

Session logs are a composite soft oracle

The source note's "session logs as composite oracle" section is useful, but there is already a broader cluster around soft oracles, automated synthesis, and quality signals. Extracting it risks making a catalogue note rather than a sharp claim. Better use it as evidence inside future notes about capture posture, promotion oracles, or behavioral faithfulness.

Grounding to reuse:

Retrieval splits by role, not content

This is a strong design move, but it may already be covered by Axes of artifact analysis, Continual learning's open problem is behaviour, not knowledge, and the source note itself. Extract only if future work repeatedly needs the exact retrieval/activation consequence of the role axis.

Grounding to reuse:

This should cite Agent memory is a crosscutting concern, not a separable niche if extracted. The retrieval split is one instance of the larger decomposition: knowledge-role access is context-engine retrieval and navigation; system-definition access is activation and behavior change.

Typed cue indexes bridge the activation gap

This exact title is still too design-specific to extract as theory. It would make a better reference or instruction artifact once fields, matching rules, priority budgets, and tests are specified. The theoretical extraction should either be System-definition cues need behavioral faithfulness tests or the broader loading-mechanism claim Typed cue indexes are on-situation context injection.

Four layers separate trace, observation, episode, and library

Do not extract as a theory note. This is the source note's architecture, not a standalone principle. If extracted, it would duplicate the synthesis without improving composability. The separable theory is narrower and better grounded as Memory layers are progressive distillation interfaces or Atomic memory and narrative memory need different retrieval interfaces.

Suggested extraction order

  1. Store-everything is a capture posture, not a memory design
  2. System-definition cues need behavioral faithfulness tests
  3. Library-derived cues should be compiled views
  4. Episodes are retrospective memory; workshops are active memory
  5. Negative results are memory artifacts because final artifacts hide discarded paths

This order fixes the source note's biggest theoretical overcommitment first, then strengthens the activation theory, then handles source-of-truth control and workshop integration. Negative results are valuable but less urgent because they do not currently create a contradiction in the architecture.

Near-term supporting candidates, not first-pass extractions:

  • Typed cue indexes are on-situation context injection - write when the context-engine loading hierarchy needs the "on situation" row.
  • Extraction taxonomy should follow oracle strength - write if the extraction taxonomy starts getting reused outside the memory-design note.
  • Memory layers are progressive distillation interfaces - write only if the four-layer architecture needs a generalized defense without preserving the exact layer count.

Impact on the source note

If these are extracted, designing-agent-memory-systems.md should become more explicitly architectural:

  • Link to Store-everything is a capture posture, not a memory design in the opening and weaken "store everything" from premise to capture substrate.
  • Link to System-definition cues need behavioral faithfulness tests in the commitment section and add a sentence that cue firing is only a candidate behavioral intervention.
  • If Typed cue indexes are on-situation context injection is not extracted, still revise the cue section to name the mechanism as on-situation loading, parallel to on-reference and on-invoke context loading.
  • Link to Library-derived cues should be compiled views in the library-to-observation backflow section and add provenance/source-hash/regeneration fields.
  • Link to Episodes are retrospective memory; workshops are active memory near the episode layer and clarify that active work surfaces are adjacent to the retrospective memory stack.
  • Link to Negative results are memory artifacts because final artifacts hide discarded paths in the negative-result use case and keep the source note focused on how the four-layer architecture stores and activates those records.
  • If Extraction taxonomy should follow oracle strength stays unextracted, still revise the extraction section to order signal types by oracle strength and promotion risk.
  • If Memory layers are progressive distillation interfaces stays unextracted, still revise the architecture section to present the four layers as one progressive-disclosure design rather than a universal layer count.

Notes not to write yet

Do not write all candidates at once. The extraction should preserve composability, not create a burst of thin notes. Start with the first two because they repair real tensions in the current synthesis. Write the others when a second artifact needs to cite them or when the source note revision would otherwise become bloated.