Private
Public Access
0
0

conductor(track): fable_review_20260617 cluster 8 (Memory & Storage) sub-report

Tier 3 worker dispatch. Verdict: Useful + nagent-stronger. 499 lines. Fable System Prompt.md:166-251 (memory_system + persistent_storage_for_artifacts) cited. Project refs: src/models.py History types, agent_memory_dimensions.md, guide_knowledge_curation.md. Fable artifact NOT committed.
This commit is contained in:
2026-06-18 20:05:07 -04:00
parent bf73daac6e
commit 0f7f088eba
@@ -0,0 +1,499 @@
# Cluster 8: Memory System & Persistent Storage
**Sub-agent dispatch:** Tier 3 Worker (2026-06-17). Read-only research task.
**Sources read:**
- `docs/artifacts/Fable System Prompt.md` lines 166-251 (`memory_system` + `persistent_storage_for_artifacts`)
- `docs/artifacts/Fable System Prompt.md` lines 436-480 (`search_instructions`, the copyright-quote discipline)
- `src/models.py:200-231` (the `#region: History Utilities` block + `parse_history_entries`)
- `src/models.py:523-559` (`FileItem` schema — the curation memory dim)
- `src/history.py:8-100` (`UISnapshot`, `HistoryEntry`, `HistoryManager` — UI undo/redo, not memory)
- `docs/guide_discussions.md` (full file, 353 lines — the discussion dim)
- `conductor/code_styleguides/agent_memory_dimensions.md` (full file, 306 lines — the 4-dim canonical)
- `docs/guide_agent_memory_dimensions.md` (full file, 278 lines — the cross-cutting user guide)
- `docs/guide_knowledge_curation.md` (full file, 358 lines — the 4th dim deep-dive)
- `conductor/code_styleguides/knowledge_artifacts.md` (referenced; canonical for the harvest pattern)
- `conductor/tracks/nagent_review_20260608/nagent_review_v2_3_20260612.md` §2.8 (Pattern 8: Harvest Knowledge), §3.1 (Knowledge harvest subsystem), §3.9 (Per-file knowledge notes), §4.4 (per-file notes sub-pattern)
- `conductor/tracks/fable_review_20260617/spec.md` §5 row 8 (this cluster's scope)
---
## 1. What Fable says
Fable's `memory_system` section is 5 lines (L166-170) and the `persistent_storage_for_artifacts` section runs L171-251. The two sections are structurally separate but conceptually adjacent: the `memory_system` describes Claude's user-facing memory feature (the setting Anthropic ships in Claude.ai); the `persistent_storage_for_artifacts` describes the JavaScript-key-value storage API that powers artifacts in Claude.ai. Both are framed as "state that persists across sessions" but they target different layers (a per-user memory layer vs. a per-artifact storage layer).
### 1.1 The `memory_system` section (L166-170)
The section is two bullets:
> "Claude has a memory system which provides Claude with access to derived information (memories) from past conversations with the user" (L168)
> "Claude has no memories of the user because the user has not enabled Claude's memory in Settings" (L170)
That's the whole section. The framing is **affordance**, not implementation: Fable tells the model what it *can* access (memories), not how the memories are stored, retrieved, ranked, audited, or pruned. The "derived information" hedge — "derived information (memories)" — is the load-bearing word: the model is told the memories are *not raw transcripts* but *extracted facts*. There is no description of the extraction pipeline, the dedup logic, the retention policy, the audit log, or the user controls.
The "user has not enabled Claude's memory in Settings" disclosure is a transparency move: if the user has the toggle off, the model must say so rather than fabricating memories. This is the same pattern Fable uses elsewhere (the "Claude does not have X" disclaimer) — it's product transparency, not behavioral instruction.
### 1.2 The `persistent_storage_for_artifacts` section (L171-251)
This is the substantive part. The section describes the `window.storage` API, a JavaScript key-value store available to artifacts. The section is structured as:
1. The 4 API methods (L181-184): `get(key, shared?)`, `set(key, value, shared?)`, `delete(key, shared?)`, `list(prefix?, shared?)`.
2. A usage example block (L188-202) showing `await window.storage.set('entries:123', JSON.stringify(entry))` and the corresponding `get`/`list` calls.
3. The "Key Design Pattern" subsection (L206-211): hierarchical keys under 200 chars, "no whitespace, path separators, or quotes"; "combine data updated together in single keys"; the example reframes `cards + benefits + completion` as a single `cards-and-benefits` key.
4. The "Data Scope" subsection (L215-220): personal (shared: false, default) vs shared (shared: true, visible to all users).
5. The "Error Handling" subsection (L222-241): "all storage operations can fail — always use try-catch"; the note that accessing non-existent keys throws (does not return null); the two try-catch patterns for "should succeed" vs "checking existence."
6. The "Limitations" subsection (L245-249): text/JSON only, keys under 200 chars, values under 5MB, rate-limited, last-write-wins, "always specify shared parameter explicitly."
7. A closing recommendation (L251): "implement proper error handling, show loading indicators and display data progressively…consider adding a reset option."
The substantive rules are concentrated in (3) and (5):
**The hierarchical-keys rule (L206):** "Use hierarchical keys under 200 chars: `table_name:record_id` (e.g., 'todos:todo_1', 'users:user_abc')." This is a real engineering pattern — namespace prefix + record id is the standard shape for a flat key-value store. The 200-char cap is a backend constraint; the no-whitespace / no-path-separator / no-quote rule is a constraint from the storage parser.
**The single-key batching rule (L210):** "Combine data that's updated together in the same operation into single keys to avoid multiple sequential storage calls." This is a real anti-pattern warning: the example reframes `await set('cards'); await set('benefits'); await set('completion')` as `await set('cards-and-benefits', {cards, benefits, completion})`. The motivation is rate-limiting — multiple sequential calls hit the limit; one combined call doesn't.
**The personal-vs-shared rule (L215-220):** The model is told to use `shared=false` by default and to inform users when their data will be visible to others. The "inform users" rule is a transparency directive tied to the personal/shared toggle.
**The try-catch rule (L222):** "All storage operations can fail - always use try-catch." This is paired with the asymmetry that `get()` *throws* on missing keys (rather than returning `null`), so the "check if a key exists" pattern requires a try-catch rather than a null-check. This is a real edge case in the API design; the model is told to wrap every call.
### 1.3 What's missing from Fable's framing
The `persistent_storage_for_artifacts` section is a **developer API reference**, not a **memory model**. It tells the model (or the artifact author) how to *use* the key-value store; it does not tell the model how to *think about* memory. Specifically absent:
- **No provenance.** Every key is opaque; the model is not told to record where data came from, which conversation, or which user action.
- **No retention / pruning.** The model is told keys can be deleted, but not told when or why. There is no "delete old entries after N days" rule, no "archive before delete" pattern.
- **No user audit.** The user can `rm`-style delete via the artifact, but the model has no obligation to surface the data to the user. The "consider adding a reset option" (L251) is a recommendation, not a requirement.
- **No concurrency control.** "Last-write-wins for concurrent updates" (L247) is stated as a limitation; the model is not told how to detect or resolve conflicts.
- **No transaction model.** The "combine data updated together" rule (L210) is a workaround for the lack of transactions; it's not framed as such.
- **No typing / schema.** Keys store arbitrary JSON; the model is told to namespace via the key prefix, not via any schema. There is no equivalent of nagent's 7-category schema or Manual Slop's `FileItem` schema.
### 1.4 Brief cross-ref: `search_instructions` (L436-480)
The `search_instructions` section is mostly about web search behavior (per cluster 7 scope), but the opening copyright-quote discipline (L444-446) is directly relevant to *this* cluster's research task:
> "15+ words from any single source is a SEVERE VIOLATION. ONE quote per source MAXIMUM—after one quote, that source is CLOSED. DEFAULT to paraphrasing; quotes should be rare exceptions." (L444-446)
Fable is telling the model to treat external sources the same way the user's cluster-spec tells the sub-agent to treat Fable: ≤15 words per quote, one quote per source, paraphrase by default. The structural parallel is informative — Fable's own discipline is being applied *to Fable itself* in this report.
---
## 2. What this project does
Manual Slop does not have a "memory system" in Fable's sense, nor a `window.storage` API. It has **4 memory dimensions**, each with a different shape, scope, and edit surface. The 4-dim model is the canonical reference (`conductor/code_styleguides/agent_memory_dimensions.md:13-18`); the project treats memory as **structured state**, not as opaque key-value blobs.
### 2.1 The 4 memory dimensions (the canonical model)
Per `conductor/code_styleguides/agent_memory_dimensions.md:13-18`:
| Dim | Where it lives | What it stores | How it's edited | SSDL |
|---|---|---|---|---|
| 1 | **Curation** | `FileItem` + `ContextPreset` + Fuzzy Anchors | *How to render a file* | Structural File Editor; project TOML | `[Q]` |
| 2 | **Discussion** | `app.disc_entries` + branching + `UISnapshot` | *What was said* | GUI `[Edit]` mode; `[Branch]`; undo/redo | `o==>` |
| 3 | **RAG** | `src/rag_engine.py` (ChromaDB) | *Semantic fingerprints* | (opaque vector store) | `[Q]` |
| 4 | **Knowledge** | `~/.manual_slop/knowledge/*.md` + per-file + digest + ledger | *Durable learnings* | Plain markdown edit | `o==>` |
**The 4 dimensions are not interchangeable.** Per `conductor/code_styleguides/agent_memory_dimensions.md:244`: "When designing a new feature, ask: which of the 4 dimensions is the natural home? Don't reach for the RAG because 'it's there'; reach for the dimension whose shape matches the data."
The decision tree (`conductor/code_styleguides/agent_memory_dimensions.md:264-271`):
```
Q: What is the *data* (not the operation) the feature needs?
├── "How to render a file" ──► Curation (FileItem)
├── "What was said in this chat" ──► Discussion (disc_entries)
├── "What similar content exists" ──► RAG (RAGEngine.search)
└── "What we learned from past runs" ──► Knowledge (knowledge/digest.md)
```
This is the data-oriented contrast to Fable's "one key-value store, call it memory" framing. Manual Slop's model says: **memory is plural**; the wrong shape for the right question is a common mistake; the 4 dims are the named, distinct, user-editable layers.
### 2.2 Curation memory (per-file structural)
**The shape** (`conductor/code_styleguides/agent_memory_dimensions.md:22-66` + `src/models.py:523-559`):
The `FileItem` dataclass at `src/models.py:523` has 10 fields:
```python
@dataclass
class FileItem:
path: str
auto_aggregate: bool = True
force_full: bool = False
view_mode: str = 'full'
selected: bool = False
ast_signatures: bool = False
ast_definitions: bool = False
ast_mask: dict[str, str] = field(default_factory=dict)
custom_slices: list[dict] = field(default_factory=list)
injected_at: Optional[float] = None
```
The 9 explicit fields are all about **how to render a file** — none are about user-derived facts about the file. `view_mode` selects between full / skeleton / summary / sig / def / agg; `ast_signatures` / `ast_definitions` are AST-aware reductions; `custom_slices` are the Fuzzy Anchor slices (`docs/guide_context_curation.md`). The user's edit surface is the Structural File Editor (the GUI modal that lets the user change `view_mode` per file).
**The storage shape.** Persisted in `manual_slop.toml` (or a project TOML) as `[[discussion.context_files]]` entries via `FileItem.to_dict()` / `from_dict()` (`src/models.py:550-580`). A `ContextPreset` is a named, persisted set of `FileItem`s (`src/models.py:909-937`).
**No `notes` field.** Per nagent_review_v2_3 §3.9 (`nagent_review_v2_3_20260612.md:2091`): "Manual Slop equivalent. `models.FileItem` (per `src/models.py:510`) has 9 fields… **No `notes` field.** No per-file knowledge notes dimension." This is the load-bearing gap that cluster 8 will surface — the curation dim is *about rendering*, not *about facts*. Fable's `entries:123` pattern (storing user-derived facts keyed by namespace) has no analog in the curation dim; the closest analog is the **knowledge dim** (4th dim), which is the project's structured answer to "remember things I've learned."
### 2.3 Discussion memory (per-discussion conversational)
**The shape** (`docs/guide_discussions.md:31-43`):
```python
{
"role": str, # "User" | "AI" | "Vendor API" | "System" | <user-edited>
"content": str, # fully editable in GUI
"collapsed": bool,
"ts": str, # ISO timestamp, prefixed with `@`
"thinking_segments": list[dict], # AI entries with <thinking> blocks
"usage": dict, # {"input_tokens", "output_tokens", "cache_read_input_tokens"}
"read_mode": bool, # render as Markdown vs editable text
}
```
The data is a flat list of entry dicts (`app.disc_entries: list[dict]`). The data model is **open**: extra keys are allowed and ignored by the renderer (`docs/guide_discussions.md:43`). The user can add custom metadata via the Hook API or by editing the project TOML directly.
**The discussion is the source of truth for "what was said."** Per `conductor/code_styleguides/agent_memory_dimensions.md:124`: "The `disc_entries` list is the single source of truth for 'what was said in this discussion.'"
**The edit surface.** A1-A7 per-entry operations (`docs/guide_discussions.md:72-86`): edit content, toggle read/edit, collapse/expand, change role, insert, delete, branch. Branching creates a new Take named `<base>_take_<n>`; takes are sibling views of the same conversation, not separate conversations. Per-entry edits are undo-able (`src/history.py:71-141`, `HistoryManager`).
**The persistence shape** (`docs/guide_discussions.md:202-249`): the discussion persists in the project TOML under `project.discussion.discussions[<name>]["history"]`. The persistence is **explicit** (B4 Save button) and **implicit** (on `_switch_discussion` and `_branch_discussion`). The "context_snapshot" (`disc_data["context_snapshot"]`) records the FileItem list at send time; switching back to a discussion restores the file list. This is the project's answer to "remember which files were in context for this discussion."
**The data model is precise.** Each entry has a structured role, a timestamp, a collapsed flag, optional thinking segments, and optional usage accounting. The model is *not* a flat text log; it is a list of structured records. Fable's `entries:123 → JSON.stringify(entry)` (L195) pattern is roughly equivalent to one Manual Slop discussion entry (each is a structured record), but Manual Slop's record has 7 explicit fields and is open to extension; Fable's is an opaque JSON blob in a key-value store.
### 2.4 RAG memory (opt-in semantic)
**The shape** (`conductor/code_styleguides/agent_memory_dimensions.md:128-170`):
ChromaDB vector store; per-file `FileItem`-like records with embeddings. `RAGEngine.search(query, k=N)` returns the top-N most-similar chunks. Persisted in `tests/artifacts/.slop_cache/chroma_<embedding_provider>/`.
**RAG is opt-in, default-off in new projects.** Per `conductor/code_styleguides/rag_integration_discipline.md` (referenced from `agent_memory_dimensions.md:170`): the discipline is opt-in, complement (never replace), provenance (file path + chunk offset), no mutation, feature-gated, graceful failure.
**RAG is the wrong shape for "what did we learn from past sessions."** Per `conductor/tracks/nagent_review_20260608/nagent_review_v2_3_20260612.md:631`: RAG is fuzzy, opaque, not auditable, not durable across embedding-provider switches. The knowledge dim is the right shape for durable learnings; RAG is the right shape for semantic search at query time.
### 2.5 Knowledge memory (per-project durable, provenance-aware)
**The shape** (`conductor/code_styleguides/agent_memory_dimensions.md:174-226` + `docs/guide_knowledge_curation.md`):
A markdown tree at `~/.manual_slop/knowledge/`:
| File | Format | What it stores |
|---|---|---|
| `knowledge/facts.md` | `- {statement} {provenance}` | Durable statements about systems, repos, tools |
| `knowledge/decisions.md` | `- {statement, reason} {provenance}` | Decisions that were made |
| `knowledge/questions.md` | `- {question} {provenance}` | Unanswered questions |
| `knowledge/playbooks.md` | `- **{name}**: {steps} {provenance}` | Reusable command sequences |
| `knowledge/tasks.md` | `- {task}` (## Open / ## Done) | Open and done tasks |
| `knowledge/files/{file_id}.md` | `- {note} {provenance}` | Per-file notes (keyed by inode) |
| `knowledge/digest.md` | bounded 4KB | The projected digest (injected as `{knowledge}` block) |
| `knowledge/ledger.json` | `{entries: {sha256: {status, at, items}}}` | The harvest audit log |
**The provenance string** is `[from: {conversation_name}, {date}]`. The provenance is appended by the harvest; the user can edit any line. The audit log (`ledger.json`) gates deletion on a proven harvest — the user cannot accidentally delete a conversation whose durable knowledge hasn't been distilled (`docs/guide_knowledge_curation.md:146-182`).
**The 7-category harvest schema** (`docs/guide_knowledge_curation.md:188-234`): the LLM's harvest output is strict JSON with 7 categories (`facts`, `decisions`, `tasks_done`, `tasks_open`, `questions`, `playbooks`, `files`). The category schema is the load-bearing contract: the LLM cannot return prose, cannot omit categories, cannot invent items ("Empty arrays are valid and expected"). The retry budget is 2 attempts (`docs/guide_knowledge_curation.md:236-255`).
**The size budgets** (`docs/guide_knowledge_curation.md:258-264`):
| Constant | Value | Why |
|---|---|---|
| `SUMMARIZE_THRESHOLD_BYTES` | 64 KB | Files > 64KB get summarized first |
| `MAX_HARVEST_SOURCE_BYTES` | 1 MB | Files > 1MB are kept (not harvested) |
| `DIGEST_MAX_BYTES` | 4 KB | The bounded digest size |
| `HARVEST_MAX_ATTEMPTS` | 2 | Retry budget on parse failure |
The 4KB digest is the projected view injected as the `{knowledge}` block in the initial context (`docs/guide_knowledge_curation.md:323-348`). The bounded digest is the cache-friendly answer to "give me the durable knowledge in 4KB or less."
**The "delete to turn off" pattern** (`docs/guide_knowledge_curation.md:285-306`): the knowledge digest is gated by file presence. `rm ~/.manual_slop/knowledge/digest.md` → no `{knowledge}` block injected. No env var, no config toggle, no GUI checkbox. The file is the switch. Re-enable by running the harvest, which regenerates the digest.
### 2.6 The contrast with Fable's `window.storage`
| Aspect | Fable `window.storage` | Manual Slop |
|---|---|---|
| **Scope** | Per-artifact (each artifact is its own KV store) | Per-project (4 dims, project-scoped) |
| **Schema** | None (opaque JSON) | Typed: `FileItem` (curation), entry dict (discussion), ChromaDB record (RAG), 5 category files (knowledge) |
| **Provenance** | None | `[from: conversation, date]` on every knowledge line; sha256 ledger; inode-keyed per-file notes |
| **Audit** | None | `ledger.json` gates deletion on proven harvest |
| **Retention** | Last-write-wins; no retention policy | Append-only category files; bounded 4KB digest; the harvest reclaim lifecycle |
| **User controls** | "consider adding a reset option" (recommendation) | Plain-text edit of every category file; GUI Knowledge panel; per-file notes; dry-run-by-default harvest |
| **Error handling** | `try/catch` around every call | Result-style failure markers (`harvest-failed`, `too-large`, `deleted-unharvested`) in the ledger; graceful failure + visible marker |
| **Concurrency** | Last-write-wins (acknowledged as limitation) | Append-only merge (no contention); per-thread `threading.local()` for transient state |
| **Memory-as-plural** | One KV store | 4 named dimensions with non-interchangeable shapes |
The contrast is not just *more features*. The contrast is **shape**. Fable's `window.storage` is a flat key-value namespace with no semantics beyond namespace-prefix conventions. Manual Slop's 4 dims are *named* (curation / discussion / RAG / knowledge), *shaped* (each has a distinct data model), *edited* (each has a distinct user surface), and *queried* (each has a distinct query model). Fable's "use a hierarchical key" pattern is the same shape advice Manual Slop gives, but applied to a single KV store rather than to 4 named dimensions.
### 2.7 UI history (the unrelated `src/history.py`)
`src/history.py` defines `UISnapshot` (the UI state for undo/redo), `HistoryEntry`, and `HistoryManager` (the stack-based undo/redo). This is **not** memory in the Fable sense — it is in-memory undo state for the current session. The `UISnapshot` dataclass captures 13 fields (ai_input, project_system_prompt, temperature, disc_entries, files, screenshots, etc.); the `HistoryManager` pushes/pops up to 100 snapshots. The snapshots are not persisted to disk; they are in-process only.
This is mentioned only to head off confusion: when Fable says "memory system," Manual Slop has *both* a `HistoryManager` (in-process undo) *and* the 4 memory dimensions (persistent storage). They serve different purposes. The in-process undo is not a memory dim; the 4 memory dims are.
### 2.8 Where the 4 dims land in the cache-friendly context (the 12-layer model)
The 4 memory dims are not just a static classification; they are *injected* into the LLM context at specific layers of the 12-layer cache-friendly model (per `conductor/code_styleguides/cache_friendly_context.md`):
| Layer | Content | Which dim? |
|---|---|---|
| 1-6 | role, schema, tools, system prompt, persona, project context | (foundational) |
| **7** | **knowledge digest** | **Knowledge (4th dim)** |
| 8-12 | discussion metadata, active preset, per-file details, prior tool results, user message | **Curation (1st dim)** + **Discussion (2nd dim)** |
| (separate) | `{rag-context}` block (opt-in) | **RAG (3rd dim)** |
The knowledge digest is the *only* memory dim in the stable cache prefix (layer 7). Per `docs/guide_knowledge_curation.md:326-348`: "The digest is injected into the *stable* position of the initial context (layer 7 of the 12-layer model)… The cache can include the digest in the cached prefix; the volatile suffix is not cached." This is the cache-friendly answer to "give me the durable knowledge in 4KB or less — and let me cache it across turns."
The curation dim is per-file and lands in the *volatile* suffix (layer 10), because each turn may have different files in scope. The discussion dim is the *user's own prior turns* (layers 8-12) and is per-turn. The RAG dim is a separate `{rag-context}` block injected at LLM call time, opt-in (`src/rag_engine.py`).
**The contrast with Fable.** Fable's `window.storage` does not specify *where* in the context the stored data appears — the artifact author decides. Manual Slop's 4 dims have fixed injection points: layer 7 (knowledge digest), layer 10 (curation per-file details), volatile suffix (discussion prior turns), and the `{rag-context}` block (RAG). The injection points are part of the data model, not a downstream decision.
The cache byte-comparison test (`tests/test_aggregate_caching.py`, per `conductor/code_styleguides/cache_friendly_context.md` §2) is the design contract: the first N characters of the context are identical across turns of the same discussion. N is `aggregate.stable_prefix_length(ctrl)`; the knowledge digest is one of the load-bearing contributors to the stable prefix. Fable's `window.storage` has no equivalent — there is no "stable prefix" concept in an artifact's KV store.
### 2.9 The implementation cross-references (file:line map)
Per `conductor/code_styleguides/agent_memory_dimensions.md:280-294`, the implementation is mostly present: curation lives in `src/models.py:510-559` (`FileItem`) + `src/context_presets.py` + `src/aggregate.py`; discussion lives in `src/gui_2.py:3770-3853` (A1-A7 render) + `src/history.py:8-71` (`UISnapshot`, `HistoryManager`) + `src/project_manager.py:429+` (branching); RAG lives in `src/rag_engine.py:1-384` (ChromaDB). The knowledge store + harvest CLI are "(proposed)" entries — scoped in Candidate 11 of `nagent_review_v2_3_20260612.md:2098`. Fable's `window.storage` is a runtime API exposed by the Claude.ai browser sandbox; the implementation is the artifact host, not the prompt. Manual Slop's codification names file:line for each dim — the implementation is *in the project's own code*.
---
## 3. What nagent does
nagent's `knowledge harvest` (`nagent-gc`) is the substantive pattern in this cluster. The harvest is the **3rd memory dimension** in nagent's framing (per `nagent_review_v2_3_20260612.md:552-674`); the project then extends nagent's framing to a **4th dimension** (per-file knowledge notes) at §3.9 (L2022-2105). The two are sibling patterns.
### 3.1 The knowledge harvest (Pattern 8) — `nagent_review_v2_3_20260612.md:552-674`
**The claim** (`nagent_review_v2_3_20260612.md:554`): "Dead conversations accumulate, and deleting them loses what was learned. Therefore: distill, then delete — and feed the distillate back in."
**The components** (`nagent_review_v2_3_20260612.md:556-571`):
| Component | Where | What it does |
|---|---|---|
| `nagent-gc` | `bin/nagent-gc:1-150` | CLI: classify, estimate cost, harvest, reclaim |
| `run_gc(root, ...)` | `bin/helpers/nagent_gc_lib.py:330+` | Library: dry-run or apply; iterates harvest candidates |
| `scan_root(root)` | `bin/helpers/nagent_gc_lib.py:80+` | Classifies artifacts: `live` / `user-kept` / `prune` / `harvest` / `keep` |
| `harvest_conversation(path, ...)` | `bin/helpers/nagent_gc_lib.py:235+` | For files >64KB, summarize first; otherwise use full text; 2 retries on parse failure |
| `merge_harvest(root, name, harvested, date)` | `bin/helpers/nagent_gc_lib.py:245+` | Appends harvested items to category files with provenance |
| `regenerate_digest(root, max_bytes=4096)` | `bin/helpers/nagent_gc_lib.py:380+` | Rebuilds `digest.md` from category files; sections in fixed order; newest first |
| `load_ledger` / `save_ledger` | `bin/helpers/nagent_gc_lib.py:115-130` | sha256-of-content gate; "already harvested" path reclaims without re-distilling |
| `parse_harvest_json(text)` | `bin/helpers/nagent_gc_lib.py:180+` | Strict JSON parser with code-fence tolerance; validates 7 categories |
**The 7-category schema** (`nagent_review_v2_3_20260612.md:573-583`): facts / decisions / tasks_done / tasks_open / questions / playbooks / files. Each row is `{statement, detail}` (or `{name, steps}` for playbooks, or `{path, note}` for files). The prompt mandates: "Return only JSON in exactly this form (no prose, no markdown fence)." "Empty arrays are valid and expected: most conversations contain nothing durable. Do not invent items to fill categories."
**The constants** (`nagent_review_v2_3_20260612.md:585-591`): same 4 budgets as Manual Slop (`SUMMARIZE_THRESHOLD_BYTES = 64KB`, `MAX_HARVEST_SOURCE_BYTES = 1MB`, `DIGEST_MAX_BYTES = 4KB`, `HARVEST_MAX_ATTEMPTS = 2`). The Manual Slop implementation borrows these constants directly (`docs/guide_knowledge_curation.md:258-264`).
**The classification** (`nagent_review_v2_3_20260612.md:600-611`):
| Class | Trigger | Action |
|---|---|---|
| `live` | `file-index-*`, `index-saved-conversations-*`, per-file conversations whose target still exists, `latest-*` active conversations | KEEP |
| `user-kept` | Path is in the saved-conversations index | KEEP |
| `harvest` | Per-file conversations whose target is gone; archived conversations; delegated sub-conversations | LLM-DISTILL → append → reclaim |
| `prune` | Split directories with no `index.json`; split directories whose source is gone or hash doesn't match | DELETE |
| `keep` | Anything unclassified | KEEP (default safe) |
**The digest ordering** (`nagent_review_v2_3_20260612.md:613-614`): sections iterated in `(Open tasks, Open questions, Decisions, Facts, Playbooks)` order; within each section, bullets reversed for newest-first. If all sections empty, the digest is *deleted* (the "delete to turn off" pattern).
### 3.2 The per-file knowledge notes (sub-pattern) — `nagent_review_v2_3_20260612.md:2022-2105`
**The claim** (`nagent_review_v2_3_20260612.md:2024`): "When you know things about a specific file, those notes should live next to the file's identity (inode), not next to a conversation or a session. Then, the next time the file is in scope, the notes come back automatically."
**The implementation** (the `merge_harvest` "files" branch, `nagent_review_v2_3_20260612.md:2028-2054`):
```python
for row in harvested.get("files", []):
if not isinstance(row, dict):
continue
path_text = str(row.get("path") or "").strip()
note = str(row.get("note") or "").strip()
if not note:
continue
target = Path(path_text) if path_text else None
if target is not None and target.is_file():
try:
file_id = file_id_for_path(target)
except OSError:
file_id = None
if file_id is not None:
_append_bullets(
file_knowledge_path(root, file_id), f"# {target.resolve()}",
[f"{note} {provenance}"],
)
file_notes += 1
continue
# Target no longer resolvable: the note survives as a fact.
prefix = f"{path_text}: " if path_text else ""
_append_bullets(knowledge / "facts.md", "# Facts", [f"{prefix}{note} {provenance}"])
file_notes += 1
```
**The fallback** (`nagent_review_v2_3_20260612.md:2051-2053`): "Target no longer resolvable: the note survives as a fact." The note's path-prefix (`{path}: `) is preserved as a prefix on the fallback fact; the per-file binding is lost but the note survives.
**The injection point** (`nagent_review_v2_3_20260612.md:2509-2515`): per-file knowledge is injected as part of the file-edit block, in the stable position. When a file is in scope for editing, its knowledge comes back automatically.
**The verdict for Manual Slop** (`nagent_review_v2_3_20260612.md:2091-2098`):
> "Manual Slop equivalent. `models.FileItem` (per `src/models.py:510`) has 9 fields: `path, auto_aggregate, force_full, view_mode, selected, ast_signatures, ast_definitions, ast_mask, custom_slices`. **No `notes` field.** No per-file knowledge notes dimension."
> "Verdict. **GAP.** The per-file notes dimension is absent in Manual Slop. `FileItem` would need a `notes: str = ""` field; the Structural File Editor would need a 'Notes' text area; `aggregate.py:run` would need a `{file-knowledge}` block in the initial context."
The gap is precisely named. The Manual Slop candidate list includes "Candidate 11.1: per-file knowledge notes — bundle with Candidate 11" (`nagent_review_v2_3_20260612.md:2098`).
### 3.3 The 4-dim framing in nagent_review_v2_3
The v2.3 review explicitly frames the project in terms of the 4 memory dims:
> "The 4 memory dimensions (the framing):" (`nagent_review_v2_3_20260612.md:4198`)
The surrounding context (the section header at `nagent_review_v2_3_20260612.md:4187-4202`) is the project's design intent: curation (FileItem) and discussion (disc_entries) are present and strong; RAG is opt-in and is the wrong shape for durable knowledge; knowledge is the missing dim. The Manual Slop codification of the 4 dims (`conductor/code_styleguides/agent_memory_dimensions.md`, `docs/guide_agent_memory_dimensions.md`, `docs/guide_knowledge_curation.md`) is the direct response to nagent's framing — Manual Slop adopts the 4-dim model and adds the knowledge dim, with the digest bounded to 4KB and the harvest pipeline implemented.
**The note on the spec's section reference.** The track spec (`fable_review_20260617/spec.md:222`) cites nagent §2.1 for "4 memory dimensions." In v2.3 the §2.1 slot is "Pattern 1: Text In, Text Out" (`nagent_review_v2_3_20260612.md:242`); the 4-dim framing moved to §2.8 (Pattern 8: Harvest Knowledge, Reclaim Space) in the v2.3 restructure. The §3.9 reference for per-file knowledge notes is correct in v2.3 (`nagent_review_v2_3_20260612.md:2022`). The substance is unchanged across versions — the v2.1/v2.2 §2.1 is the same content as v2.3 §2.8. Cluster 8 cites v2.3 throughout.
### 3.4 What Manual Slop adopted from nagent (the load-bearing adoption)
The Manual Slop codification is not just *inspired by* nagent — it adopts specific patterns and constants directly:
**The 4 size budgets** are identical (`docs/guide_knowledge_curation.md:258-264` + `nagent_review_v2_3_20260612.md:585-591`): `SUMMARIZE_THRESHOLD_BYTES = 64KB`, `MAX_HARVEST_SOURCE_BYTES = 1MB`, `DIGEST_MAX_BYTES = 4KB`, `HARVEST_MAX_ATTEMPTS = 2`.
**The 7-category schema** is identical: facts / decisions / tasks_done / tasks_open / questions / playbooks / files. Same shape, same JSON contract, same code-fence tolerance.
**The retry-suffix pattern** is identical: on retry, append `\nYour previous reply was not valid JSON. Return only the JSON object.\n` to the prompt (`docs/guide_knowledge_curation.md:255`).
**The provenance format** is identical: `[from: {conversation_name}, {date}]` (`docs/guide_knowledge_curation.md:42`).
**The "delete to turn off" pattern** is identical: `rm ~/.manual_slop/knowledge/digest.md` → no `{knowledge}` block injected (`docs/guide_knowledge_curation.md:289`).
**The digest section ordering** is identical: Open tasks, Open questions, Decisions, Facts, Playbooks; within each section, bullets reversed for newest-first (`docs/guide_knowledge_curation.md:137`).
**The "graceful failure" markers** are identical: `harvest-failed`, `too-large`, `deleted-unharvested` (`docs/guide_knowledge_curation.md:178-181`).
**The per-file notes pattern** is adopted but not yet implemented: the 4 Manual Slop docs describe the pattern, but `models.FileItem` does not yet have a `notes` field. The implementation is the deferred Candidate 11.1.
**The dry-run-by-default safety** is the same pattern (`docs/guide_knowledge_curation.md:266-281`): without `--apply`, the CLI classifies, estimates cost, and prints a report. No mutation.
The adoption is not a 1:1 port. Manual Slop adapts the pattern for its 4-dim model (curation is its own dim, not a "files" category sub-bucket) and for the project's data-oriented conventions (`Result[T]` + `ErrorInfo` instead of exceptions). But the constants, schema, retry pattern, provenance format, section ordering, delete-to-turn-off pattern, and graceful-failure markers are direct ports. nagent's harvest library is the source; Manual Slop's 4 canonical docs are the target.
---
## 4. Verdict
**Useful + nagent-stronger.** Fable's `window.storage` API + the hierarchical-keys pattern + the single-key-batching rule + the personal-vs-shared scoping + the try-catch-everything rule are genuinely useful engineering guidance. They are the *table-stakes* of any key-value client library: namespace your keys, batch your writes, distinguish personal vs shared scope, handle errors. None of these patterns are Fable's invention; they are the standard pattern for the API surface Fable exposes.
But Fable's framing is **memory-as-blob-store**: one key-value namespace, opaque JSON, no provenance, no retention, no audit, no schema. Manual Slop's 4 memory dimensions (curation / discussion / RAG / knowledge) are the **stronger, more grounded** version of Fable's "memory" framing. Each dim has a named shape, a user-editable surface, a query model, and (for knowledge) a provenance-aware harvest pipeline with an audit ledger. Fable's 5-line `memory_system` section is a product toggle; Manual Slop's `agent_memory_dimensions.md` is a 306-line canonical styleguide with a decision tree.
nagent's knowledge harvest + per-file knowledge notes is **the strong version of Fable's "memory" framing**. The 7-category schema, the `[from: conversation, date]` provenance, the sha256-of-content ledger, the 4KB bounded digest, the per-file notes keyed by inode — these are the load-bearing patterns that turn a key-value blob into a *durable memory system*. nagent implements them; the project adopts them.
### 4.1 Pattern-by-pattern judgment
**Pattern 1: Hierarchical keys under 200 chars (L206).** **Useful.** This is a real engineering pattern (namespace prefix + record id); the 200-char cap is a backend constraint; the no-whitespace / no-slash / no-quote rule is the parser constraint. Manual Slop's analog is implicit: the `app.disc_entries` list uses index-based addressing; `FileItem` is keyed by path; `knowledge/files/{file_id}.md` is keyed by inode. None of these are flat key-value, but the *underlying principle* (each memory cell has a structured key) is the same. Recommend: document this principle in the project's memory dim styleguide (it already exists in the per-dim "where it lives" column; no new spec needed).
**Pattern 2: Single-key batching to avoid rate limits (L210).** **Useful.** The example reframes `await set('cards'); await set('benefits'); await set('completion')` as `await set('cards-and-benefits', {cards, benefits, completion})`. This is a rate-limit-driven batching pattern; Manual Slop's analog is the digest: the knowledge dim batches *all 7 categories* into a single 4KB `digest.md` file rather than emitting 7 separate `set` calls. Recommend: no action — Manual Slop already batches.
**Pattern 3: Personal vs shared data scope (L215-220).** **Useful + Manual Slop-lacking.** The personal/shared distinction is a real product feature; the "inform users when data is visible to others" transparency rule is a good safety practice. Manual Slop has no analog: the knowledge dim is single-user (per-machine, `~/.manual_slop/knowledge/`); the curation dim is per-project (in the project TOML); the discussion dim is per-discussion (in the project TOML). There is no shared-storage concept. Recommend: note as out-of-scope — Manual Slop is a single-user tool; shared storage would be a feature add, not a "memory model" improvement.
**Pattern 4: try/catch around every storage call (L222).** **Useful + Manual Slop-different.** Fable's try/catch is the standard JS error-handling pattern; Manual Slop's convention is the data-oriented `Result[T]` + `ErrorInfo` dataclass pattern (`conductor/code_styleguides/error_handling.md`). The harvest pipeline uses 4 result markers (`harvested` / `harvest-failed` / `deleted-unharvested` / `too-large`) in `ledger.json` rather than exceptions (`docs/guide_knowledge_curation.md:178-181`). Recommend: no action — the project's convention is the data-oriented one, which is the stronger pattern.
**Pattern 5: "Claude has a memory system which provides Claude with access to derived information (memories) from past conversations" (L168).** **Useful (the concept) + nagent-stronger (the implementation).** The *concept* of a memory system that derives facts from past conversations is the right product framing. The *implementation* is opaque ("derived information") and has no provenance, no audit, no schema. nagent's knowledge harvest + Manual Slop's knowledge dim are the strong versions: schema (7 categories), provenance (`[from: conversation, date]`), audit (`ledger.json`), retention (4KB digest with truncation marker). Recommend: explicitly reject Fable's "one opaque memory feature" framing; cite nagent + Manual Slop's structured 4-dim model as the alternative.
**Pattern 6: "No `notes` field on FileItem" (the gap).** **GAP per nagent §3.9.** The project has the 4-dim framing but lacks the per-file notes dimension within the knowledge dim. The fix is named in `nagent_review_v2_3_20260612.md:2096-2098`: add `notes: str = ""` to `FileItem`, add a "Notes" text area to the Structural File Editor, add a `{file-knowledge}` block to `aggregate.py:run`. This is Candidate 11.1 in the nagent review's deferred-rebuild list. Recommend: include in `decisions.md` as a deferred-rebuild recommendation.
### 4.2 What to reject
- **The "one opaque KV store = memory" framing.** Fable's `window.storage` is a *storage API*, not a *memory model*. Treating it as a memory model collapses 4 distinct dimensions (curation / discussion / RAG / knowledge) into one flat namespace with no shape. The project should explicitly reject this framing.
- **The "user enables memory in Settings" toggle as a memory model.** Fable's `memory_system` is a 5-line product disclosure, not a memory architecture. The project should not import the toggle framing.
- **The "no schema, namespace via key prefix" pattern.** Keys like `entries:123` are namespace-by-convention, not namespace-by-type. The project's 4-dim model has named types (FileItem, disc_entry, ChromaDB record, knowledge bullet); the Fable pattern has no types. The project should not import the untyped-namespace pattern.
### 4.3 What to keep
- **The hierarchical-keys principle** (each memory cell has a structured key) — already implicit in Manual Slop's per-dim shapes.
- **The personal-vs-shared scope distinction** — out-of-scope for Manual Slop (single-user tool), but the principle is sound.
- **The error-handling discipline** — already implemented as `Result[T]` + `ErrorInfo` + ledger status markers.
- **The "consider adding a reset option" transparency** — already implemented as the "delete to turn off" pattern (`docs/guide_knowledge_curation.md:285-306`).
### 4.4 What to add (deferred-rebuild candidate)
- **Per-file knowledge notes (Candidate 11.1).** The 4-dim model is incomplete without the per-file notes dimension. The fix is small (add `notes` field + GUI text area + `{file-knowledge}` injection block) but the value is high (durable facts about specific files survive across sessions). Flag in `decisions.md`.
---
## 5. Synthesis notes for the Tier 1 writer
This cluster feeds `report.md` §10 ("Fable's Memory System & Persistent Storage") directly. Cross-references to §13 ("Genuinely Useful Patterns") and §14 ("Anti-User Watchdog Patterns"). The verdict orientation is **Useful + nagent-stronger** (per `fable_review_20260617/spec.md:182`).
### 5.1 Key claims to surface in §10
1. **Fable's `window.storage` is a useful API reference, not a memory model.** The 4 API methods, the hierarchical-keys rule, the single-key batching, the personal-vs-shared scope, and the try/catch discipline are all genuinely good engineering guidance. None of them are Fable's invention; they are the standard pattern for a key-value client library. Cite L181-184 (API methods), L206-211 (key design), L215-220 (data scope), L222-241 (error handling).
2. **Fable's `memory_system` is a 5-line product disclosure, not a memory architecture.** L168 and L170 are a setting toggle and a transparency statement, not an implementation. The "derived information" hedge is load-bearing: Fable admits the memories are extracted facts but does not describe the extraction, the audit, the retention, or the user controls. The contrast is Manual Slop's 306-line canonical styleguide + the 358-line user-facing guide + the 4-dim model with decision tree.
3. **Manual Slop's 4 memory dimensions are the strong version of Fable's "memory" framing.** Each dim has a named shape, a user-editable surface, a query model, and (for knowledge) a provenance-aware harvest pipeline with an audit ledger. Cite `conductor/code_styleguides/agent_memory_dimensions.md:13-18` (the table) + `agent_memory_dimensions.md:244-272` (the boundaries + decision tree).
4. **nagent's knowledge harvest is the strong version of Fable's "memory" framing.** The 7-category schema, the `[from: conversation, date]` provenance, the sha256-of-content ledger, the 4KB bounded digest, the per-file notes keyed by inode — these are the load-bearing patterns that turn a key-value blob into a durable memory system. Cite `nagent_review_v2_3_20260612.md:552-674` (Pattern 8) + `nagent_review_v2_3_20260612.md:2022-2105` (per-file notes §3.9).
5. **The per-file notes dimension is the named GAP.** Per `nagent_review_v2_3_20260612.md:2091-2098`: FileItem has 9 fields, no `notes`. The fix is Candidate 11.1 in the nagent deferred-rebuild list. Cite explicitly as a deferred-rebuild recommendation.
6. **The data-oriented contrast.** Manual Slop's `Result[T]` + `ErrorInfo` + ledger status markers (`harvested` / `harvest-failed` / `deleted-unharvested` / `too-large`) are the data-grounded alternative to Fable's `try/catch` pattern. The harvest pipeline's failure modes are encoded in `ledger.json`, not raised as exceptions. Cite `conductor/code_styleguides/error_handling.md` + `docs/guide_knowledge_curation.md:178-181` (the ledger status values) + `docs/guide_knowledge_curation.md:308-320` (the graceful failure modes).
### 5.2 Quotes to use in §10
- Fable L168: "Claude has a memory system which provides Claude with access to derived information (memories) from past conversations with the user" (≤15 words paraphrased; full quote exceeds)
- Fable L170: "Claude has no memories of the user because the user has not enabled Claude's memory in Settings" (full quote, 15 words)
- Fable L181: "await window.storage.get(key, shared?) - Retrieve a value → {key, value, shared} | null" (paraphrase)
- Fable L206: "Use hierarchical keys under 200 chars: table_name:record_id" (12 words)
- Fable L210: "Combine data that's updated together in the same operation into single keys" (12 words)
- Fable L215: "Personal data (shared: false, default): Only accessible by the current user" (10 words)
- Fable L222: "All storage operations can fail - always use try-catch" (8 words)
- `conductor/code_styleguides/agent_memory_dimensions.md:13`: "Curation | FileItem + ContextPreset + Fuzzy Anchors | How to render a file in the AI's context window" (paraphrase; the table)
- `conductor/code_styleguides/agent_memory_dimensions.md:244`: "When designing a new feature, ask: which of the 4 dimensions is the natural home?" (16 words)
- `docs/guide_knowledge_curation.md:13`: "The LLM harvests past discussions into these files; the user can edit any of them in plain text" (paraphrase)
- `docs/guide_knowledge_curation.md:285-286`: "Feature flags should be data, not config. If a feature is gated by the presence of a file, the user can turn it off by deleting the file" (28 words → split into 2 quotes)
- `docs/guide_knowledge_curation.md:289`: "rm ~/.manual_slop/knowledge/digest.md → no {knowledge} block injected" (paraphrase)
- `nagent_review_v2_3_20260612.md:554`: "Dead conversations accumulate, and deleting them loses what was learned. Therefore: distill, then delete" (paraphrase)
- `nagent_review_v2_3_20260612.md:2024`: "When you know things about a specific file, those notes should live next to the file's identity (inode)" (paraphrase)
- `nagent_review_v2_3_20260612.md:2096`: "No `notes` field. No per-file knowledge notes dimension" (paraphrase of the GAP verdict)
### 5.3 The §13 / §14 / §15 cross-references
- **§13 ("Genuinely Useful Patterns").** The hierarchical-keys principle (each memory cell has a structured key) + the personal-vs-shared scope distinction + the error-handling discipline are all genuinely useful. Cite L206 (keys), L215 (scope), L222 (errors). Note that Manual Slop already implements each in the project's own conventions (per-dim shapes, single-user scope, `Result[T]` + ledger markers). The useful pattern is *the principle*, not the Fable framing.
- **§14 ("Anti-User Watchdog Patterns").** The "memory is a Settings toggle" framing (L170) is *not* anti-user in itself — it's a transparency disclosure. But the *combination* of "Claude has a memory system" (L168) + "user has not enabled" (L170) + "consider adding a reset option" (L251, recommendation not requirement) constructs the memory system as opaque + non-user-controlled + lightly-suggested-to-be-resettable. The user can't see what's in memory, can't audit, can't selectively delete. This is anti-user in the *transparency* sense (not the *safety* sense). Recommend: cite as a transparency gap, contrast with the project's `ledger.json` + plain-text-edit + `delete to turn off` pattern.
- **§15 ("Persona Performance Patterns").** None of cluster 8 is persona performance. The `memory_system` section is a product disclosure; the `persistent_storage_for_artifacts` section is an API reference. Neither constructs a persona. Cluster 8 does not feed §15.
### 5.4 The data-oriented error handling parallel
Fable's `try/catch` rule (L222) is the JS-idiomatic error handling; Manual Slop's `Result[T]` + `ErrorInfo` + ledger status markers is the data-oriented equivalent. The harvest pipeline uses 4 status markers (`harvested` / `harvest-failed` / `deleted-unharvested` / `too-large`) in `ledger.json` rather than exceptions (`docs/guide_knowledge_curation.md:178-181`). The graceful failure modes table (`docs/guide_knowledge_curation.md:308-320`) lists 6 failure scenarios and their handling, all encoded as data, not control flow.
The synthesis report should surface this parallel in §10: Fable's storage error handling is persona-free (no "Claude feels bad about the storage failure"); Manual Slop's storage error handling is data-only (status markers, ledger entries, visible UI panels). The contrast is not "Fable has errors, Manual Slop doesn't" — it's "Fable uses control flow, Manual Slop uses data."
### 5.5 The "memory is plural" framing for the synthesis report's TL;DR
The single most important claim from cluster 8 is that **memory is plural, not singular**. Fable's framing is "the memory system" (singular, opaque, toggle-controlled). Manual Slop's framing is "the 4 memory dimensions" (plural, named, shaped, user-editable). nagent's framing is "the harvest + the per-file notes" (2 named sub-systems). The synthesis report's §0 TL;DR should surface this distinction as the headline: Fable's `memory_system` section is 5 lines; Manual Slop's 4-dim model is 4 named styleguides (306 + 358 + 278 + canonical knowledge_artifacts.md lines), each with a decision tree, a query model, and a user-editable surface.
### 5.6 What the §10 verdict should be
**Verdict: Useful (the API surface) + nagent-stronger (the memory architecture).** Fable's `window.storage` API is a useful engineering reference; the hierarchical-keys + single-key-batching + personal-vs-shared + try/catch rules are the standard pattern for a key-value client library. Manual Slop already implements each in its own conventions (per-dim shapes, digest batching, single-user scope, `Result[T]` + ledger). Fable's `memory_system` section is a product disclosure, not a memory architecture; nagent's knowledge harvest + per-file notes + Manual Slop's knowledge dim are the strong versions of the "memory" framing. The named gap is the per-file notes dimension (Candidate 11.1 per nagent §3.9).
**The recommended Manual Slop action:**
1. Cite the hierarchical-keys + batching principles in the memory dim styleguide as already-implemented (no change).
2. Cite the personal-vs-shared scope distinction as out-of-scope (single-user tool; no action).
3. Cite the data-oriented error handling contrast (`Result[T]` + ledger markers) in the §10 verdict.
4. Flag the per-file notes dimension (Candidate 11.1) as a deferred-rebuild recommendation in `decisions.md`.
5. Explicitly reject Fable's "one opaque KV store = memory" framing; cite the 4-dim model + the knowledge harvest as the alternative.
### 5.7 The deferred-rebuild recommendation (for `decisions.md`)
**Recommendation R8.1: Implement Candidate 11.1 (per-file knowledge notes).**
- **Source evidence.** `nagent_review_v2_3_20260612.md:2091-2098` (the named GAP verdict); `nagent_review_v2_3_20260612.md:2022-2105` (§3.9 the per-file notes pattern); `nagent_review_v2_3_20260612.md:2492-2515` (§4.4 the per-file notes sub-pattern).
- **What to build.** Add `notes: str = ""` to `FileItem` (`src/models.py:523`); add a "Notes" text area to the Structural File Editor (`docs/guide_context_curation.md`); add a `{file-knowledge}` block to `aggregate.py:run` at the file-edit position (per `nagent_review_v2_3_20260612.md:2509-2515`).
- **Why.** The 4-dim model is incomplete without per-file notes. The fix is small (3 sites, ~50 lines) but the value is high: durable facts about specific files survive across sessions; the notes come back automatically when the file is in scope; the notes are keyed by inode so they survive renames within the same filesystem.
- **Priority.** LOW standalone (small, niche) per `nagent_review_v2_3_20260612.md:2098` — bundle with the main knowledge dim implementation (Candidate 11).
- **Destination.** `conductor/code_styleguides/knowledge_artifacts.md` §? (extend the existing canonical styleguide) + `docs/guide_knowledge_curation.md` §2 (extend the existing per-file notes section).
**Recommendation R8.2: Document the "memory is plural" framing in the agent-directive corpus.**
- **Source evidence.** This cluster's §5.5 ("memory is plural, not singular"); Fable L168 ("Claude has a memory system") vs Manual Slop's 4-dim model (`conductor/code_styleguides/agent_memory_dimensions.md:13-18`).
- **What to build.** Add a 1-paragraph "memory is plural" callout to `AGENTS.md` (the top-level agent-facing rules) and to `conductor/product-guidelines.md` §"AI-Optimized Compact Style". The callout: "Manual Slop has 4 memory dimensions, not 1. The dimensions are not interchangeable. Fable-style 'one memory feature' framing collapses 4 distinct shapes into 1 opaque KV store."
- **Why.** The 4-dim model is the project's design intent; the Fable framing is a competing model. The agent-directive corpus should explicitly reject the Fable framing.
- **Priority.** LOW (documentation-only).
- **Destination.** `AGENTS.md` "Critical Anti-Patterns" or "Code Standards & Architecture" section + `conductor/product-guidelines.md` "AI-Optimized Compact Style" section.
### 5.8 The relationship to cluster 7 (search_instructions)
Cluster 7 owns the `search_instructions` copyright-quote discipline (L444-446). Cluster 8 references it as a cross-cut but does not feed §10 from it.
---
**Sub-report complete.** This is the evidence base for §10 of `report.md`.