Filesystem-Based Memory for LLM Agents: Organization, Evolution, and Sustainability

Formalization and study design

• The paper formalizes filesystem memory as management, search, and execution roles sharing one store for declarative memories and reusable skills.It evaluates conversational question answering and embodied task success while varying memory shape, stream scale, harness, and model strengths.Source: §1 · pp. 2–3

Organization

• Management agents grow subject-based trees, but organization reflects model style more than scale and can silently condense content unless preservation is enforced.As stores grow, hierarchy may relocate between folders, files, and in-file headings rather than consistently becoming more distributed.Source: §1 · p. 2

Value of organization

• Organization reliably reduces search cost as material grows, but no memory shape wins correctness everywhere and no measured agent converts organization into better answers.For skills, verbatim episode logs favor strong execution agents, whereas distilled guidance favors weak ones.Source: §1 · p. 3

Backbone model capability

• Search-agent capability improves conversation answers directly, whereas management capability mainly changes organizational style and matters for skill distillation after a threshold.A stronger management backbone recovers about half the loss from inconsistently recording changed preferences as dated updates on one benchmark, but has no effect where search already succeeds.Source: §1 · p. 3

Sustainability under scaling

• Store health remains stable within measured horizons, while taxonomy adherence erodes with growth except for the strongest management agent.Conversational stores create few files early and edit without deleting; skill stores preserve early memories, while curation effort does not become cheaper per episode.Source: §1 · p. 3

Harness

• Changing the tool set reshapes memory organization without changing outcomes, making the harness a control knob whose direction and payoff depend on the setting.Long-dialogue stores shard and tie, while skill stores consolidate and benefit.Source: §1 · p. 3

2.1 THE MEMORY STORE

• A memory store is a finite set of described markdown files in a rooted path hierarchy whose folders, filenames, and headings form a navigable taxonomy.Files expose names and one-line descriptions before their contents; optional frontmatter can carry additional metadata.Source: §2.1 · pp. 4–5
• Well-organized stores require sibling distinction and relatedness, parent-child coverage, proximity of related content, and structural economy that aids search.These five principles apply throughout the tree, including headings, and are operationalized by the paper’s hierarchy metrics.Source: §2.1 · p. 5

2.2 TOOL HARNESS AND AGENT ROLES

• A tool harness H mediates every agent interaction with the store, ranging from sandboxed shells and memory-tool functions to regex- or keyword-search extensions.Harnesses differ in granularity and power, making the same store afford different behavior under different H.Source: §2.2 · p. 5
• Management integrates each incoming chunk into a store trajectory, while search returns cited answers from a fixed store and execution uses retrieved skills to attempt tasks.Search is read-only in intent, and execution outcomes provide the probe connecting store quality to task success.Source: §2.2 · pp. 5–6

2.3 TWO INSTANTIATIONS, ONE STORE

• Conversational memory streams dialogue slices into stores queried for answers, whereas procedural memory distills task trajectories into reusable skills under a leak-free sequence.In the skill setting, later-task performance measures transfer from stores built only from earlier tasks.Source: §2.3 · p. 6
• Both instantiations share the same store class, role contracts, and harnesses, differing only in chunk and query types so one system supports declarative and procedural memory.Dialogue slices serve conversational memory, while rendered trajectories and tasks serve skills.Source: §2.3 · p. 6

2.4 WHAT IS HELD FIXED AND WHAT VARIES

• Experiments hold the execution agent and role contracts fixed while varying the store builder, stream scale, harness, and management and search policy strengths.Measurements track downstream quality, building and search cost, and store health over the store trajectory.Source: §2.4 · pp. 6–7

Benchmarks

• The conversational evaluation spans LoCoMo, PersonaMem at 32k and 128k tokens, and REALTALK, covering long multi-session, persona-rich, and naturalistic dialogue.LoCoMo uses 158 questions across multi-hop, temporal, open-domain, and single-hop categories; PersonaMem uses exact-match grading, and REALTALK contains 85 questions over 21 messaging days.Source: §3 · p. 7

Memory variants

• Six conversational memory variants range from Closed-book and chunk retrieval to verbatim, foldered, reorganized, and agent-curated stores.Reorganized store is tested in condense and preserve versions to separate layout changes from lossy versus content-preserving rewriting; agent-curated stores vary content and organization together.Source: §3 · p. 7

Models and prompts

• The main management and search agents use gpt-5.4-mini at high reasoning effort, while a fixed calibrated gpt-5.4-mini judge grades conversational correctness and citation support.Search prompts are held within one shared prompt family, require answers grounded in and cited to the store, and adapt only to store form.Source: §3 · p. 7

Tool sets

• The management agent writes through generic memory-tool file operations, whereas the search agent navigates a read-only filesystem-native harness, with Center, Center+BM25, and Shell compared in the harness study.This asymmetric tool design exposes differences in names, layout, and text; it is varied experimentally rather than treated as a finding.Source: §3 · p. 8

Controlled studies

• Strength, harness, and scale studies change one factor at a time, varying builder and searcher models, tool sets, and PersonaMem’s 32k versus 128k tiers while holding the remaining setup fixed.The searcher study forms a three-by-three model grid, and the skill experiments separately vary curator strength, execution strength, and one long chain versus three shorter chains.Source: §3 · p. 8

Metrics and skill setting

• Evaluation records correctness, citation attribution, task success, token and tool costs, intrinsic compute bounds, and store structure; ALFWorld adds five memory variants under three-chain and full-chain protocols.ALFWorld uses 140 held-out household tasks across six goal families, including no store, episode-log, curated-skills, curated-skills+memory, and guidance-synthesis variants.Source: §3 · pp. 8–9

4.1 WHAT STORES AGENTS BUILD (RQ1)

• Management agents build radically different hierarchies: conversational stores relocate structure into headings as scale grows, while procedural stores consolidate monotonically with stronger builders.Conversational builders produced 122, 2, and 105 files across the model ladder, versus 114, 45, and 16 procedural files; organization is therefore model- and setting-dependent rather than governed by one scale law.Source: §4.1 · pp. 10–12

4.2 MEMORY SHAPE AND RETRIEVAL (RQ2)

• Organization reliably buys search economy on large stores, but no memory shape wins correctness everywhere across conversational and skill benchmarks.PersonaMem reorganizers cost 1.4–1.6 cents per query versus 4.0–3.9 cents for verbatim dumps, while skill-store winners invert with execution-agent strength.Source: §4.2 · pp. 13–14

4.3 BACKBONE MODEL CAPABILITY (RQ3)

• Management-model strength changes organization more than conversational answer quality, whereas stronger search models improve reading and gpt-5.4 management enables procedural distillation.Conversational correctness ranged from 73.8% to 66.7% to 71.4% across nano, mini, and gpt-5.4 builders, while search correctness rose from about 62% to 71% to 79%; in skills, gpt-5.4 added 13 points over the lower management models.Source: §4.3 · pp. 15–17

4.4 SUSTAINABILITY AS MEMORY GROWS (RQ4)

• Conversational stores grow mainly by editing existing files: file counts stabilize early, sections grow linearly, build effort per chunk stays flat, and final size ranges from 35% above input to one quarter of input.LoCoMo elaborates through structure, references, and locators, whereas PersonaMem 128k compresses; neither organization nor growth guarantees a smaller store.Source: §4.4 · p. 17, Figure 5 · p. 18
• 51.4% to 91.4%: in a 140-task chain, every store variant improved without degradation, with the weakest execution tier gaining most as invalid actions fell from 40% to 24%.The mini-tier Episode log rose from 51.4% over its first 35 tasks to 91.4% over its last 35, exceeding the stronger tier’s three-chain average.Source: §4.4 · p. 18
• Accumulation benefits most when execution is weak: Episode log gains +7.1 points there versus +1.4 with strong execution, though differing protocols make the deltas suggestive.The no-store control is unchanged, while GS is nearly protocol-insensitive; the result reproduces under reversed end-to-end accumulation, with nano rising from 63% to 89%.Source: §4.4 · pp. 18–19
• Curated stores stabilize while episode logs expand, making retrieval costlier: GS costs $2.97 versus $7.88 for Episode log over 140 tasks.The gpt-5.4 store has 14 of its final 16 files by task 35, whereas the mini Episode log reaches 140 files and 270 kilobytes; Figure 7 tracks the growing retrieval gap.Source: §4.4 · p. 19, Figure 7 · p. 19
• Early memories overwhelmingly survive, while stronger management shifts from creating files to editing them in place; taxonomy adherence remains stable only for the strongest manager.The gpt-5.4 manager’s creations flatline by a quarter of the chain, edits continue climbing, and it leaves one of fourteen early files untouched versus nano’s twenty-two of thirty-five.Source: §4.4 · p. 20, Figure 8 · p. 20

4.5 HARNESS EFFECTS (RQ5)

• Tool choice alone radically changes store shape: Shell shards PersonaMem 128k into 147 files, whereas file-tool sets consolidate it into one or two richly sectioned files.On LoCoMo, store shape varies more across runs for Center+BM25, but the directional contrast remains: Shell partitions dense material while file-native editing favors fewer, larger files.Source: §4.5 · p. 21, Table 10 · p. 21
• Despite divergent layouts, tool sets produce near-tied conversational quality, with LoCoMo at 86.1–86.7% correctness and PersonaMem 128k at 61.9–66.7%.The search agent adapts its strategy to layout, using BM25 routinely on many-file stores but usually navigating a table of contents in a mega-file.Source: Table 9 · p. 20, §4.5 · pp. 21–22
• On skills, adding ranked search changes behavior without outcomes, whereas Shell improves the mini chain to 82.9%, ahead of Center+BM25 by +11 and Center by +9.Shell also produces a more consolidated mini store—36 versus 45 files—while preserving structure and schema; across settings, harness effects depend on whether browsing or procedural execution is rewarded.Source: §4.5 · p. 22

5 CONCLUSION

• Organization buys search economy for large content, but no single memory shape consistently wins answer quality across agents and settings.Verbatim episodes lead with strong execution agents but invert with weak ones; only Curated skills+mem (GS) survives that capability drop.Source: §5 · p. 22
• No measured agent converted organization itself into better answers; outcomes instead moved with writing capability, memory form, and curation tools.Changing a tool altered behavior without changing outcomes, whereas replacing the tool set reshaped stores through sharding, tying, consolidation, or winning.Source: §5 · pp. 22–23

A PROMPTS

• The appendix reproduces the experiment prompts, including management, source-attribution, search, closed-book, reorganization, skill-setting, and judging prompts.Each prompt is reproduced as sent, with reasoning-only think-tag scaffolding removed; tool schemas were supplied separately through named tool profiles.Source: A PROMPTS · p. 25

A.1 MANAGEMENT AGENT (BUILDER) PROMPT

• The management prompt builds curated stores incrementally with seven write tools, integrates chunks into existing memory, preserves attribution, and repairs duplication, contradiction, scattering, and stale structure.The prompt treats the filesystem as a search-oriented taxonomy whose labels, hierarchy, naming, cross-references, temporal context, and source locators support retrieval.Source: §A.1 · pp. 25–28

A.2 REORGANIZER PROMPT (REORGANIZED STORE)

• The reorganizer rewrites a verbatim store in place for up to five passes, reusing the management taxonomy and attribution guidance with preserve-every-fact and lossy condense variants.The two variants differ in one strategy rule: Reorg. (preserve) explicitly keeps every fact, whereas Reorg. (condense) leaves the model’s lossy default.Source: §A.2 · p. 28
• The reorganizer surveys content first, preserves every factual detail and source locator, and restructures files, folders, headings, and cross-references to maintain a coherent taxonomy.It corrects duplication, contradiction, scattering, stale metadata, broken references, and temporal ambiguity while keeping files readable and names informative.Source: §A.2 · p. 30
• Reorganized stores use one topic per file, YAML frontmatter, descriptive names, full-path cross-references, resolved dates, historical context, self-contained facts, and inline source attribution.Frontmatter must remain consistent with filenames and contents, while references must be updated after moves, renames, or deletions.Source: §A.2 · pp. 30–31

A.3 SEARCH AGENT PROMPTS (ONE PER MEMORY VARIANT)

• Search prompts are quality-matched across memory variants, adapting routing and evidence instructions to hierarchical files, raw session folders, verbatim dumps, and chunk retrieval.Filesystem variants use read-only view, grep, table-of-contents, and section-read tools, whereas chunk retrieval uses ranked chunk retrieval alone; all variants require cited factual answers.Source: §A.3 · p. 31
• Search agents survey once, route by names and descriptions, probe with scoped distinctive-token grep, read only needed content, verify every query component, and stop when evidence suffices.They widen searches when scoped probes miss, inspect all plausible files for open-ended completeness, distinguish multiple-choice evidence from keyword hits, and report absence only after whole-tree checks.Source: §A.3 · p. 32
• Answers must lead with the direct response, cite every factual claim using memory paths and line or section references, and state uncertainty or genuine absence explicitly.For multiple-choice questions, agents must choose an option based on supported claims and query polarity rather than merely matching option terms.Source: §A.3 · pp. 32–33
• The search prompts prioritize targeted retrieval, complete inspection for open-ended queries, evidence verification, citation, and stopping once every part of a question is supported.Filesystem and verbatim-dump variants use grep, scoped reads, and source locators; chunk retrieval uses iterative BM25 queries and cited chunk files.Source: §A.3 · pp. 35, 37
• Multiple-choice prompts treat distinctive option terms as search cues, require judging each option’s claim and polarity, and mandate choosing the option most consistent with evidence.They prohibit treating keyword hits as proof and require uncertainty notes when evidence is thin.Source: §A.3 · p. 37
• Closed-book evaluation uses an empty store and one tool round, requiring direct answers from reasoning or world knowledge without fabricated citations and with honest uncertainty.Multiple-choice responses must still name exactly one option, while unanswerable open questions may be declined briefly.Source: §A.3 · p. 38

A.4 JUDGE PROMPTS (ANSWER GRADING)

• Frozen judge prompts score answer correctness against the gold answer and assess citation correctness, relevance, and coherent aggregation from resolved cited-file contents.The correctness judge assigns 1 when the response sufficiently answers consistently with the ground truth and 0 otherwise.Source: §A.4 · p. 38
• A.4 grades responses on citation correctness, citation relevance, and coherent aggregation, using 0, 0.5, or 1.0 ratings.The judge checks whether citations support claims, whether cited information is necessary, and whether sources are synthesized coherently.Source: §A.4 · p. 39

A.5 SKILL-SETTING PROMPTS

• Curators judge outcomes before distilling trajectories, create positive procedures only from successes, encode failures as lessons or corrections, avoid duplicates, and continually reshape stale or conflicting structure.Updates should generalize demonstrated methods without inventing unsupported steps, while maintenance merges, relocates, splits, renames, or removes entries as needed.Source: §A.5 · pp. 40, 42
• A.5 requires skills and notes to share a taxonomy whose labels, scope, proximity, and ordering make retrieval efficient rather than merely increasing storage.Context-specific notes belong beside the material they annotate, while ordered steps are explicitly numbered and independent siblings remain unordered.Source: §A.5 · pp. 40–41
• The retriever receives a persistent markdown library, writes guidance verbatim for an execution agent that cannot open files, and appends provenance lines for each used source.Citations count only when the guidance uses an entry’s content, and each file should be cited at most once.Source: §A.5 · pp. 43–44
• The retriever’s cost model charges every turn for the entire accumulated conversation, motivating batched independent probes, avoiding repeated reads, and stopping when further rounds no longer change the guidance.The prompt recommends surveying the filesystem once and widening scoped searches only when initial probes miss relevant material.Source: §A.5 · pp. 43–44
• The skill retriever surveys, routes, probes, reads, and composes task-specific guidance from relevant skills, warnings, observations, and transferable procedures.It should gather potentially relevant entries before judging what to include, then prioritize warnings and operational facts while avoiding unsupported invention.Source: §A.5 · p. 44
• Retriever guidance must restate transferable advice for the current task, phrase variable conditions conditionally, preserve exact-object discipline, and include recorded locations and action-enabling facts.The response should contain only useful, library-grounded content; if no experience applies, it must return the prescribed no-experience sentence.Source: §A.5 · p. 44
• The execution agent is fixed across conditions: its prompt embeds the task and retrieved guidance, while each step supplies the current observation and admissible actions for selecting the next action.The required response format encloses reasoning in <think> tags and the chosen admissible action in <action> tags.Source: §A.5 · p. 45

Store overview

• Across the examples, stores vary from three subject files to deep consolidation and flat procedural shelves, depending on the stream and management agent.The snapshots include a small conversational store, two shapes for the same PersonaMem stream, and a final skill store.Source: B EXAMPLE MEMORY FILESYSTEMS · p. 45

LoCoMo

• The LoCoMo store consolidates one conversation into three speaker-and-thread files with 116 sections and 151 cross-references.The files sit under people/, with shared content linked rather than duplicated.Source: B EXAMPLE MEMORY FILESYSTEMS · pp. 45–46

PersonaMem Foldered

• Foldering the same PersonaMem stream creates 10 topic directories containing 20 unchanged session files, with topical rather than temporal grouping.Several directories contain only one session, and the contents remain byte-identical to the Verbatim dump.Source: B EXAMPLE MEMORY FILESYSTEMS · p. 46

Skill store

• The strongest management agent’s skill store contains 16 files with no directories: 5 skills, 11 notes, and 101KB organized by imperative filenames.The five skills cover goal families, while the notes record preconditions, search strategy, and environment quirks; weaker models produce 114 and 45 mostly flat files.Source: B EXAMPLE MEMORY FILESYSTEMS · pp. 46–47

C.1 CONFIGURATION

• Every main-comparison role uses gpt-5.4-mini, and all reported cells completed fully on test splits.This configuration covers the search agent, store-building management agents, and judge; calibration and pilots used validation splits, while a mid-run API failure invalidated an entire cell.Source: §C.1 · p. 47, Table 11 · p. 48
• Build streams contain consecutive dialogue chunks of at most eight turns or 3,000 characters, while session-based stores retain one whole session per file.Management integrates one chunk per build episode; Chunk retrieval indexes those chunks, whereas Verbatim dump and Foldered sessions use whole sessions.Source: §C.1 · p. 47
• Agent inputs and outputs are not truncated during normal operation, with caps of 8192 tokens for search and judging, 32768 for curation, and 100k for store restructuring.The caps prevent runaway generations rather than constrain observed runs, and file views are never clipped.Source: §C.1 · pp. 47–48
• Truncation was zero across the 2,885-turn validation scan and every reported cell except one benign foldered-store search event.The exception occurred on a single foldered-store query.Source: §C.1 · p. 48
• Contexts exceeding 96k prompt tokens are compacted into a running summary plus the three most recent turns, with compaction logged for every cell.The same compaction rule applies throughout the experiments.Source: §C.1 · p. 48
• Build episodes cap at 60 tool rounds, while search episodes cap at 40 rounds for curated and reorganized stores and 20 for other variants.The 60- and 40-round limits were never reached; the deepest observed episodes used 36 build and 24 search rounds, while one Chunk retrieval LoCoMo query reached 20 rounds and was unchanged at 40.Source: §C.1 · p. 48
• Agents use the provider’s high-effort setting without overrides, while the judge samples at temperature 0.Closed-book answers use one search round, and the only observed 20-round cap hit was re-probed without changing the answer.Source: §C.1 · p. 48

C.2 BENCHMARK DATA PROVENANCE

• LoCoMo uses held-out conversation conv-50, evaluating 158 non-adversarial questions from its 204-question test set.Source: §C.2 · p. 48
• The remaining benchmarks use one REALTALK conversation with 85 questions, three PersonaMem 32k conversations totaling 32 questions, and one PersonaMem 128k conversation with 42 questions.The PersonaMem 128k case contains a 784-message prefix spanning 20 sessions and about 113k tokens.Source: §C.2 · p. 48

C.3 FIXED SEARCH PROMPTS

• Each memory variant uses a fixed search-agent prompt, while Reorganized and Agent-curated stores share the hierarchical-store prompt.Full prompt texts and variant assignments are provided in Section A.Source: §C.3 · p. 49

C.4 TOOL DESCRIPTIONS AND SCHEMAS

• Tools are delivered through function calling: management uses seven write tools, search uses four read-only tools, and variants add ranked whole-file search, chunk retrieval, or shell access.Center+BM25 adds ranked whole-file search, Shell replaces both tool sets with bash, and Foldered sessions uses view, grep, and rename.Source: §C.4 · p. 49, Table 12 · p. 49
• The management interface supports constrained insertion, deletion, and renaming within /memories, including frontmatter and extension safeguards.Insertions cannot occur inside YAML frontmatter; recursive deletion excludes the root, and renaming creates missing parent directories while enforcing path rules.Source: §C.4 · p. 50
• Search tools differ materially: BM25 searches whole files recursively, chunk retrieval returns the top three full-text chunks, and shell exposes an isolated filesystem with standard command-line tools.Whole-file BM25 searches filenames, frontmatter, and bodies; chunk retrieval is the only memory-access method for its agent.Source: §C.4 · p. 50, §C.5 · p. 50

C.5 COST ACCOUNTING

• Cost accounting logs uncached-input, cached-input, and output tokens per model call, allowing pricing to be applied while dollar figures retain run-to-run variation.Main-grid costs use gpt-5.4-mini rates of $0.75, $0.075, and $4.50 per million uncached-input, cached-input, and output tokens, respectively.Source: §C.5 · pp. 50–51
• Intrinsic efficiency compares measured billing with no-cache and perfect transcript-level bounds, and the implementation was validated by exactly reproducing two published cells.The perfect bound pre-fills every transcript token once and decodes every output once; the efficiency share is (no-cache − measured)/(no-cache − perfect).Source: §C.5 · p. 51

C.6 SKILL-SETTING CONFIGURATION

• The skill setting evaluates 140 ALFWorld valid-seen tasks across six goal families using reset three-chain and fixed-order family-interleaved full-chain streams, with identical full-chain order across cells.No-store full chains reproduce their three-chain counterparts task for task, indicating chain order and machinery contribute nothing independently.Source: §C.6 · p. 51
• Every reported skill cell completed without errored tasks, while gated stores passed file-by-file schema and outcome-gate audits, including the nano management agent.Costs separate deployment from build and use single-run dollar measurements under the Section C.5 token framework.Source: §C.6 · p. 51
• The instruction-corrected Curated skills+mem variant was validated but deliberately not run, so its effect on that condition remains an open limitation.The corrected management-agent instruction was not tested in the reported matrix.Source: §C.6 · p. 51

D.1 WHY THE AGENT-CURATED STORE FAILS ON PERSONAMEM 32K

• 37.5% correctness makes the agent-curated store the weakest PersonaMem 32k store, versus 78.1% for the verbatim dump on identical questions.Agent curation also requires the most tool calls across benchmarks.Source: §D.1 · pp. 51–52
• The failure is representational rather than coverage or reliability: all thirteen gold-critical facts missed by curation remain present, while searches find and cite the right sections.Cells have no API errors, failed episodes, or truncation; grounding scores are 0.94–0.98 and attribution scores are 0.97–1.0.Source: §D.1 · p. 51
• Curation leaves superseded preferences as current traits, flattening temporal distinctions that PersonaMem questions use to separate latest states from earlier or static information.For example, a current forum preference conflicts with a later record of joining and enjoying a forum.Source: §D.1 · pp. 51–52
• Curation also neutralizes first-person affect and scatters narrative arcs across files, making appreciation and decisive updates harder to distinguish or read together.An emphatic statement that an app became a game-changer is rewritten as a capability list.Source: §D.1 · p. 52
• Correctness declines with distance from chronological records: Verbatim dump 78.1 > Foldered sessions 62.5 > Agent-curated store 37.5 on the same questions.The graded damage links increasingly transformed store structure to poorer performance.Source: §D.1 · p. 52
• Raising the management agent from gpt-5.4-mini to gpt-5.4 improves fixed-search-agent correctness from 37.5 to 56.3, supporting a build-side capability effect but not closing the gap.The upgrade recovers 18/32 versus 12/32 questions; the paired result gains nine and loses three, with one-sided sign test p=0.073.Source: §D.1 · p. 52
• The proposed explanation remains unproven: management inconsistently reconciles dated updates, search sometimes misses or mishandles conflicts, and the search-side test remains outstanding.The same management upgrade is flat on the 128k conversation, while 56.3 remains below the verbatim dump’s 78.1.Source: §D.1 · p. 52

• The audit identified 4 of 158 LoCoMo golds as materially defective, spanning unsupported entities, invented attributes, absent answers, and speaker-memory misattribution.The defects occur in questions 20, 64, 112, and 138; evaluation still uses the standard dataset.Source: §D.2 · p. 52
• Removing four defective questions leaves orderings unchanged and raises every non-closed-book variant by roughly a point; the unfiltered table remains the headline result.Table 13 uses the same per-question judge outcomes after filtering, while the standard data remains primary for comparability.Source: §D.2 · p. 52
• On question 138, memory-equipped variants answered faithfully from the transcript but were marked wrong, whereas closed-book guessing alone matched the defective gold.Source: §D.2 · p. 52

Build-side effort

• Build effort followed store shape more than harness: the 29-file LoCoMo Center+BM25 store required 15.1 rounds and 25.4 tool calls per chunk, versus 14.8 rounds and 20.0 calls for the 39-file shell store.The single-file Center+BM25 PersonaMem build was cheapest at 9.6 rounds per chunk, compared with 11.6 for shell; Center dollar figures are reference-only because they came from a separate run.Source: §D.3 · pp. 52–53

Re-run provenance

• Re-running corrected two prompt defects from scratch found null-to-negligible behavioral impact, while the weighted search-score composite shifted by +4.0, −2.7, +0.6, and +2.8 points across paired cells.One wording-only correction changed a repeated build from 2 files to 29, motivating caution about interpreting single-run store shapes.Source: §D.3 · p. 53

D.4 HIERARCHY METRICS: DEFINITIONS AND FULL PANEL

• The panel measures tree shape and taxonomy adherence using lexical TF-IDF metrics, covering depth, fanout, concentration, references, sibling labels, cohesion, scope leakage, and tree-distance alignment.Shape metrics include mean and maximum leaf depth; B1–B4 quantify sibling label distinguishability, sibling-content cohesion, scope leakage, and Spearman correlation between tree and content distances.Source: §D.4 · p. 53
• Search-serving structure is evaluated through retrieval effort, while hierarchy metrics describe contract adherence rather than intrinsic memory quality.The section explicitly leaves the relationship between adherence and retrieval value open.Source: §D.4 · p. 53
• B4 tends to increase with finer granularity because many small topical files give tree distance more signal than consolidated files organized mainly by headings.Table 14 applies the full hierarchy panel, including shape and taxonomy-contract metrics, across stores.Source: §D.4 · p. 53, Table 14 · p. 54
• B3 detects content-level misplacement but not label-level misnesting, motivating a parent-child label-consistency check as a refinement.Source: §D.4 · pp. 53–54

D.5 SKILL SETTING: SUCCESS BY GOAL FAMILY

• At gpt-4.1-mini, Curated skills+mem (GS) leads the Episode log in five of six families, including two-object (+20.8), cool (+20.0), heat (+18.8), examine (+15.4), and placement (+2.8).The Episode log retains only clean, where it leads by +7.4.Source: §D.5 · p. 54
• At gpt-4.1, the Episode log leads through clean (85.2) and cool (84.0), while Curated skills+mem (GS) leads heat (87.5 versus 37.5).Heat and cool remain the weak execution agent’s hardest families, with heat never exceeding 37.5.Source: §D.5 · p. 54
• Simple placement is near ceiling for every variant at both tiers, ranging from 94.3 to 100, so harder families determine the matrix.The micro column reproduces Table 5 exactly.Source: §D.5 · p. 54

D.6 SKILL SETTING: DEPLOYMENT COST BY ROLE AND ABSOLUTE COMPUTE

• Curated skills+mem (GS) is the cheapest variant per solved task at both tiers, costing 5.7 versus 6.1 cents at the mini tier.At gpt-4.1-mini, its retrieval costs $2.97 versus $7.88 for Episode log; execution dominates gpt-4.1 costs, while curation costs $7–11 per cell.Source: §D.6 · p. 55
• At gpt-4.1-mini, Curated skills+mem (GS) reduces episodes to 23.6 versus 32.0 rounds and lowers uncached and cached input to 13.5k and 135k tokens.Episode log most shortens gpt-4.1 episodes, but at mini its verbatim context raises cached input and output above the no-store row.Source: §D.6 · p. 55
• The naive no-cache sum overstates intrinsic cost 16 to 23 times, while appending episodes capture 94.4 to 96.3 percent of achievable caching savings.Measured compute remains within 1.7 to 2.05 times the physical minimum.Source: §D.6 · p. 55
• At the mini tier, Curated skills+mem (GS) beats Episode log on every intrinsic-compute component, including first-turn context of 620 versus 2,684 tokens.Its roughly 29 percent intrinsic advantage is robust to prefill-versus-decode re-weighting; at gpt-4.1, Episode log is instead cheapest at 6,937 tokens per task.Source: §D.6 · p. 55

D.7 GROWTH ATLAS

• Every full-chain task records outcomes, per-role token usage, and store snapshots, with conversational builds retaining equivalent per-chunk records.These records support the growth figures and enable cross-checking store reconstructions against per-task snapshots.Source: §D.7 · p. 55
• Figure 9 measures content-level hierarchy across 560 validated states with zero mismatches and tracks adherence erosion, while Figure 10 measures skill-side compression.The conversational counterpart to Figure 10 is the bytes panel of Figure 5.Source: §D.7 · p. 55
• Figure 11 measures curation effort that does not amortize, supporting the six-to-twelve-round growth finding, while Figure 12 separates running-success curves by goal family.Figure 11 uses per-episode curation records; Figure 12 reports differenced curves.Source: §D.7 · p. 55

• AI assistants generated illustration assets for Figures 1 and 2, while every plot and table reports measured experimental results rather than AI output.Source: E USE OF AI ASSISTANTS FOR… · p. 55