docs: align section numbers with file numbers across all design docs

Renumber all headings so file 0N uses section N.x consistently:
- 03: 4.x → 3.4-3.6
- 04: 5.x → 4.x, # Graph-LLM → ## 4.6 with ### subsections
- 05: 6/7/8 → 5.1/5.2/5.3, remove sub-numbering on ###
- 06: 9/10 → 6.1-6.5, remove sub-numbering on ###
- 07: 11.x → 7.x
- 08: 12/13 → 8.1/8.2
- ZH: fix ### → ## heading levels for 01/02/06/07

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Author: Grivn

docs/design/03-concepts.md

@@ -93,7 +93,7 @@ oplog (
 
 ---
 
-## 4. System Architecture
+## 3.4 System Architecture
 
 Mnemon's architecture is divided into five layers:
 
@@ -178,7 +178,7 @@ mnemon/
 └── Makefile                   # Build, install, test
 ```
 
-## 4.1 Data Directory Layout
+## 3.5 Data Directory Layout
 
 ```
 ~/.mnemon/
@@ -197,7 +197,7 @@ mnemon/
 
 **Isolation boundary**: Each store contains an independent `mnemon.db` — insights, edges, and oplog are fully isolated. Prompt files (`guide.md`, `skill.md`) are shared — behavioral rules are universal, memory data is private.
 
-## 4.2 Store Isolation
+## 3.6 Store Isolation
 
 Mnemon supports named stores for lightweight data isolation between different agents, projects, or scenarios.
 

docs/design/04-graph-model.md

@@ -10,7 +10,7 @@ Mnemon implements four graphs, each capturing one dimension of relationships:
 
 ![MAGMA Four-Graph Model](../diagrams/04-magma-four-graph.jpg)
 
-## 5.1 Temporal Graph
+## 4.1 Temporal Graph
 
 **Purpose**: Capture the chronological order of memories, building a temporal skeleton of the knowledge flow.
 
@@ -30,7 +30,7 @@ Insight A (2h ago) ←── backbone ──→ Insight B (1h ago) ←── bac
 
 **Metadata**: `{"sub_type": "backbone"|"proximity", "hours_diff": "2.34"}`
 
-## 5.2 Entity Graph
+## 4.2 Entity Graph
 
 **Purpose**: Link insights that mention the same entities.
 
@@ -50,7 +50,7 @@ Insight A ←── entity ──→ Insight B ←── entity ──→ Insigh
 
 **Metadata**: `{"entity": "Qdrant"}`
 
-## 5.3 Causal Graph
+## 4.3 Causal Graph
 
 **Purpose**: Capture the reasons behind decisions and cause-effect relationships.
 
@@ -73,7 +73,7 @@ Insight A ──── causal ────→ Insight B
 
 This is a quintessential example of the LLM-Supervised philosophy: Binary handles low-cost candidate discovery (regex + token overlap), while the LLM handles high-value causal judgment.
 
-## 5.4 Semantic Graph
+## 4.4 Semantic Graph
 
 **Purpose**: Connect semantically similar insights based on meaning.
 
@@ -94,7 +94,7 @@ Insight C ←── semantic (LLM review) ──→ Insight D
                 cos=0.65, manually linked after LLM judged "related"
 ```
 
-## 5.5 Four-Graph Synergy: Intent-Adaptive Weighting
+## 4.5 Four-Graph Synergy: Intent-Adaptive Weighting
 
 Different query intents activate different graph traversal weights:
 
@@ -109,11 +109,11 @@ When asking "why was SQLite chosen," the causal edge weight is highest, so the s
 
 ---
 
-# Graph-LLM Theoretical Foundations
+## 4.6 Graph-LLM Theoretical Foundations
 
 The following sections establish the theoretical basis for why graph databases are the native storage model for LLMs, and why `remember / link / recall` constitutes a universal protocol for agent memory systems.
 
-## Structural Isomorphism
+### Structural Isomorphism
 
 LLM attention, graph data models, and natural language all describe the same thing: weighted associations between entities.
 
@@ -125,7 +125,7 @@ Natural Language:  subject ←predicate→ object
 
 Relational databases force network relationships into tables. Vector databases retain only one relationship type (similarity). Only graphs preserve full relational semantics.
 
-## The Three-Step Paradigm: Extract → Candidate → Associate
+### The Three-Step Paradigm: Extract → Candidate → Associate
 
 Graph construction engines universally decompose into three steps:
 
@@ -147,7 +147,7 @@ The three-step model is a spectrum — the more semantically rich the data model
 | **Vector** | Text → embedding | ANN dedup | Metadata only (single relation type) |
 | **KV** | Key:value | Key existence check | _(nearly none)_ |
 
-## Read-Write Symmetry (Unique to Graphs)
+### Read-Write Symmetry (Unique to Graphs)
 
 On graph databases, the read and write paths mirror each other using the same three-step model:
 
@@ -169,7 +169,7 @@ This symmetry does NOT hold for other database types — relational write is sch
 
 **Implication**: An LLM needs to master only one cognitive pattern to handle both graph reads and writes.
 
-## From the LLM Perspective: Query → Reason
+### From the LLM Perspective: Query → Reason
 
 Regardless of the underlying database, LLM interactions on the read side collapse to two steps:
 
@@ -183,7 +183,7 @@ This is the RAG paradigm applied to any data store. The variation lies in the tr
 - **Text-to-Cypher**: must understand graph structure
 - **Text-to-Vector**: encode only, near-zero translation
 
-## Other Storage Types as Degenerate Graphs
+### Other Storage Types as Degenerate Graphs
 
 | Storage Type | What's Lost Compared to Graph |
 |-------------|------------------------------|
@@ -194,7 +194,7 @@ This is the RAG paradigm applied to any data store. The variation lies in the tr
 
 A vector database can answer "what is **similar** to what" but cannot answer "what **caused** what" or "what **belongs to** what". Graphs can.
 
-## remember / link / recall as Universal Algebra
+### remember / link / recall as Universal Algebra
 
 The three-step paradigm (Extract → Candidate → Associate) maps directly to three primitive operations: **remember**, **link**, **recall**. This is not an implementation detail of mnemon — it is the minimal complete interface for any agent memory system.
 
@@ -238,7 +238,7 @@ Native RAG      link absent entirely
 
 The more degenerate the `link` operation, the more burden falls on the LLM at recall time to infer associations that were never stored.
 
-## The Protocol Gap: LLM ↔ Database
+### The Protocol Gap: LLM ↔ Database
 
 ### The Missing Layer
 
@@ -349,7 +349,7 @@ MemGPT──┘                         │── SQLite adapter (mnemon current
 - **Not competing with Mem0** on product features (it is a product bound to its implementation)
 - **Analogous to MCP** — MCP connected LLMs to the tool ecosystem; this protocol connects LLMs to the database ecosystem
 
-## Academic Landscape and Positioning
+### Academic Landscape and Positioning
 
 ### Prior Art Assessment
 
@@ -401,7 +401,7 @@ Original contributions (no prior formulation found)
 - Transformers are Graph Neural Networks (arXiv 2506.22084, Jun 2025)
 - A Generalization of Transformer Networks to Graphs (arXiv 2012.09699, 2020)
 
-## Validation: mnemon Architecture
+### Validation: mnemon Architecture
 
 mnemon's design directly reflects these insights:
 
@@ -413,7 +413,7 @@ recall   → Extract + Candidate + Associate (intent detection → multi-signal
 
 Five edge types preserve five distinct relational semantics. Degenerating to pure vector retrieval would retain only `semantic` — losing ~80% of relational information. MAGMA ablation studies confirm: removing causal edges drops accuracy 3-5%, removing temporal edges drops it further.
 
-## Summary
+### Summary
 
 - **Extract → Candidate → Associate** is the universal paradigm for graph construction engines
 - This three-step model achieves its **most complete expression** on graphs and degenerates toward KV

docs/design/05-pipelines.md

@@ -4,13 +4,13 @@
 
 ---
 
-## 6. Write Pipeline: Remember
+## 5.1 Write Pipeline: Remember
 
 `mnemon remember` is the core command for writing memories. It includes a built-in diff step that automatically detects duplicates and conflicts before storage. The write transaction executes atomically within a single SQLite transaction.
 
 ![Remember Pipeline](../diagrams/02-remember-pipeline.jpg)
 
-### 6.1 Detailed Flow
+### Detailed Flow
 
 ```
 mnemon remember "Chose Qdrant as the vector database" \
@@ -84,13 +84,13 @@ After receiving this output, the LLM can evaluate candidates and establish edges
 
 ---
 
-## 7. Read Pipeline: Smart Recall (Default)
+## 5.2 Read Pipeline: Smart Recall
 
 `mnemon recall` is Mnemon's core retrieval algorithm. Smart recall is the default mode for all queries. It combines intent detection, multi-signal anchor selection, Beam Search graph traversal, and multi-factor re-ranking to achieve intent-aware graph-enhanced retrieval. Use `--basic` for legacy SQL LIKE fallback.
 
 ![Smart Recall Pipeline](../diagrams/03-smart-recall-pipeline.jpg)
 
-### 7.1 Step 1: Intent Detection
+### Step 1: Intent Detection
 
 Query intent is automatically identified via regex matching:
 
@@ -103,7 +103,7 @@ Query intent is automatically identified via regex matching:
 
 Supports the `--intent` flag to manually override automatic detection.
 
-### 7.2 Step 2: Multi-Signal Anchor Selection (RRF Fusion)
+### Step 2: Multi-Signal Anchor Selection (RRF Fusion)
 
 Multiple signals run in parallel and are merged via Reciprocal Rank Fusion:
 
@@ -119,7 +119,7 @@ RRF Score = Σ  1 / (k + rank_i + 1)    (k = 60)
 
 Each insight may rank differently across signals; RRF fusion produces a robust composite ranking.
 
-### 7.3 Step 3: Beam Search Graph Traversal
+### Step 3: Beam Search Graph Traversal
 
 Starting from each anchor, Beam Search is performed across the four graphs:
 
@@ -153,7 +153,7 @@ for each anchor:
 
 WHY queries use a wider beam and deeper traversal because causal chains typically span multiple hops.
 
-### 7.4 Step 4: Multi-Factor Re-Ranking
+### Step 4: Multi-Factor Re-Ranking
 
 For all collected candidates, a four-dimensional score is computed and combined via weighted sum:
 
@@ -175,11 +175,11 @@ Weights vary by intent:
 | ENTITY | 0.20 | **0.40** | 0.20 | 0.20 |
 | GENERAL | 0.25 | 0.25 | 0.25 | 0.25 |
 
-### 7.5 Step 5: WHY Post-Processing — Causal Topological Sort
+### Step 5: WHY Post-Processing — Causal Topological Sort
 
 If the intent is WHY, an additional topological sort using Kahn's algorithm is performed: results are arranged along causal edges so that **causes come first, effects follow**.
 
-### 7.6 Signal Transparency
+### Signal Transparency
 
 Each retrieval result includes a detailed signal breakdown:
 
@@ -202,7 +202,7 @@ This is a unique innovation in Mnemon: **exposing the retrieval pipeline's inter
 
 ---
 
-## 8. Deduplication & Conflict Detection: Diff
+## 5.3 Deduplication & Conflict Detection: Diff
 
 ![Diff & Dedup Pipeline](../diagrams/07-diff-dedup-pipeline.jpg)
 
@@ -221,7 +221,7 @@ When `remember` is called, the built-in diff runs before the transaction:
 
 The `--no-diff` flag disables this check for cases where the caller wants unconditional insertion.
 
-### 8.1 Typical Workflow
+### Typical Workflow
 
 A single `remember` call handles everything:
 

docs/design/06-lifecycle.md

@@ -8,7 +8,7 @@ Mnemon is not an append-only system. Effective memory management requires import
 
 ![Lifecycle & Retention](../diagrams/06-lifecycle-retention.jpg)
 
-## 9.1 Effective Importance (EI)
+## 6.1 Effective Importance (EI)
 
 EI combines base importance, access frequency, time decay, and graph connectivity:
 
@@ -27,13 +27,13 @@ Interpretation:
 - **Long period without access** -> exponential decay (halves every 30 days)
 - **Rich graph connections** -> indicates relevance to other knowledge, bonus applied
 
-## 9.2 Immunity Rules
+## 6.2 Immunity Rules
 
 The following insights are exempt from automatic cleanup:
 - `importance >= 4` (high-value memories)
 - `access_count >= 3` (frequently retrieved)
 
-## 9.3 Auto-Pruning
+## 6.3 Auto-Pruning
 
 Triggered when the total number of active insights exceeds **1000**:
 
@@ -43,7 +43,7 @@ Triggered when the total number of active insights exceeds **1000**:
 4. Soft-delete (set `deleted_at`)
 5. Cascade-delete related edges
 
-## 9.4 GC Command
+## 6.4 GC Command
 
 Manual lifecycle management tool:
 
@@ -57,11 +57,11 @@ mnemon gc --keep <id>
 
 ---
 
-## 10. Embedding Support
+## 6.5 Embedding Support
 
 Embedding vectors are an optional enhancement. Without embeddings, Mnemon operates entirely on keywords and graph structure; with embeddings, semantic retrieval capabilities are significantly enhanced.
 
-### 10.1 Ollama Integration
+### Ollama Integration
 
 Via the local Ollama service (no external API required):
 
@@ -75,11 +75,11 @@ Mnemon ──HTTP──→ Ollama (localhost:11434)
 - **Graceful degradation**: Automatically falls back to token overlap when Ollama is unavailable
 - **Zero new dependencies**: Pure stdlib `net/http`
 
-### 10.2 Vector Storage
+### Vector Storage
 
 Vectors are serialized as little-endian float64 BLOBs stored in the `insights.embedding` column (768 x 8 = 6144 bytes/insight).
 
-### 10.3 Usage Scenarios
+### Usage Scenarios
 
 | Scenario | Without Embedding | With Embedding |
 |----------|------------------|----------------|
@@ -88,7 +88,7 @@ Vectors are serialized as little-endian float64 BLOBs stored in the `insights.em
 | recall -> traversal | Pure structural score | Structural + semantic similarity |
 | recall -> re-ranking | KW + Entity + Graph | KW + Entity + Similarity + Graph |
 
-### 10.4 Management Commands
+### Management Commands
 
 ```bash
 ollama pull nomic-embed-text    # Install the model

docs/design/07-integration.md

@@ -8,7 +8,7 @@
 
 Mnemon integrates with LLM CLIs through lifecycle hooks, a skill file, and a behavioral guide. Claude Code's [hook system](https://docs.anthropic.com/en/docs/claude-code/hooks) is the reference implementation — all components are deployed automatically via `mnemon setup`.
 
-## 11.1 Integration Architecture
+## 7.1 Integration Architecture
 
 Four hooks drive the memory lifecycle:
 
@@ -46,7 +46,7 @@ Three layers work together:
 | **Skill** | `SKILL.md` — command reference in Claude Code skill format | `.claude/skills/mnemon/` | Teaches the LLM *how* to use mnemon commands |
 | **Guide** | `guide.md` — detailed execution manual for recall, remember, and delegation | `~/.mnemon/prompt/` | Teaches the LLM *when* to recall, *what* to remember, and *how* to delegate |
 
-## 11.2 Hook Details
+## 7.2 Hook Details
 
 Claude Code fires hooks at specific lifecycle events. Mnemon registers up to four, each with a distinct role in the memory lifecycle:
 
@@ -97,7 +97,7 @@ Fires before context window compression. Instructs the agent to extract the most
 echo "[mnemon] Context compaction starting. Review this session and remember the most valuable insights (up to 5) before context is compressed. Delegate to Task sub-agents now."
 ```
 
-## 11.3 Automated Setup
+## 7.3 Automated Setup
 
 `mnemon setup` handles all deployment automatically:
 
@@ -142,7 +142,7 @@ Key setup options:
 
 The Prime hook is always installed. Remind, Nudge, and Compact hooks are optional (Remind and Nudge enabled by default).
 
-## 11.4 Sub-Agent Delegation
+## 7.4 Sub-Agent Delegation
 
 Memory writes don't happen in the main conversation. Instead, the host LLM delegates to a lightweight sub-agent:
 
@@ -174,7 +174,7 @@ This separation means:
 - **Context isolation**: Memory processing doesn't pollute the main conversation context
 - **Model efficiency**: Sonnet handles routine execution while Opus focuses on high-level decisions
 
-## 11.5 Adapting to Other LLM CLIs
+## 7.5 Adapting to Other LLM CLIs
 
 For CLIs with hook support, replicate the Claude Code pattern: register lifecycle hooks that call mnemon commands, deploy the skill file, and provide the behavioral guide.
 

docs/design/08-decisions.md

@@ -4,7 +4,7 @@
 
 ---
 
-## 12. Design Decisions & Trade-offs
+## 8.1 Design Decisions & Trade-offs
 
 ### Why LLM-Supervised Instead of an Embedded LLM?
 
@@ -51,7 +51,7 @@ Mnemon retains MAGMA's **architectural skeleton** (four-graph separation, intent
 
 ---
 
-## 13. Future Direction
+## 8.2 Future Direction
 
 The [two-layer architecture](02-philosophy.md#23-memory-gateway-protocol-not-database) has achieved agent-side pluggability — any LLM CLI can interact with Mnemon through the protocol surface today. The remaining work is on the other side.