REM phase: write proposed-changes.md with skill patches

docs/CLI.md

@@ -189,7 +189,7 @@ Run the dream consolidation cycle, which performs automated memory maintenance i
 
 - **Light phase:** Sort and deduplicate near-duplicate memories (cosine similarity ≥ `dream.similarity_threshold`).
 - **Deep phase:** Score, promote, decay, and merge memories. Also promotes inbox items to long-term memory (items with attention_score ≥ `dream.min_score` become permanent; lower-scored items are evicted). Decays confidence on idle memories. Boosts frequently accessed memories. Performs LLM-assisted merging of similar pairs. Auto-links memories with high cosine similarity (≥ `dream.auto_link_threshold`, default 0.85).
-- **REM phase:** Extract themes from memory clusters and write a dream diary (read-only reflection). When Ollama is available, generates actionable behavioral insights via LLM; falls back to count-based summaries when Ollama is unavailable.
+- **REM phase:** Extract themes from memory clusters and write a dream diary (read-only reflection). When Ollama is available, generates actionable behavioral insights via LLM with "Do" directives, "Verify" steps, and `[SKILL PATCH]` sections (Detection Rule, Checklist, Pitfall, Verification); falls back to count-based summaries when Ollama is unavailable. When run with `--apply`, also appends behavioral insight and skill patch sections to `proposed-changes.md` at `~/.config/llmem/proposed-changes.md` (or `LMEM_HOME/proposed-changes.md`). Each dream run's entries are separated by a timestamp header. The file is append-only — existing content is preserved.
 
 Without `--apply`, the dream cycle runs as a **dry run** — output is prefixed with `[DRY RUN]` and no changes are written to the database.
 

docs/CONFIGURATION.md

@@ -45,6 +45,7 @@ dream:
   boost_amount: 0.05
   diary_path: null             # Auto-resolved from GetDreamDiaryPath()
   report_path: null            # Auto-resolved from GetDreamReportPath()
+  proposed_changes_path: null  # Auto-resolved from GetProposedChangesPath()
   behavioral_threshold: 3
   behavioral_lookback_days: 30
   ollama_url: http://localhost:11434  # Ollama API URL for LLM-generated behavioral insights
@@ -62,7 +63,7 @@ session:
   debounce_seconds: 30        # Idle debounce interval in seconds
 ```
 
-> **Note:** The following fields exist in the Python config but are **not wired** in the Go implementation: `min_score`, `min_recall_count`, `min_unique_queries`, `boost_on_promote`, `merge_model`, `calibration_enabled`, `calibration_lookback_days`, `inbox_capacity`, `correction_detection` (top-level), `copilot` (top-level), and `proposed_changes_path`. Setting these in `config.yaml` has no effect when using the Go CLI.
+> **Note:** The following fields exist in the Python config but are **not wired** in the Go implementation: `min_score`, `min_recall_count`, `min_unique_queries`, `boost_on_promote`, `merge_model`, `calibration_enabled`, `calibration_lookback_days`, `inbox_capacity`, `correction_detection` (top-level), and `copilot` (top-level). Setting these in `config.yaml` has no effect when using the Go CLI.
 
 > **Note:** The Go config resolves `db_path` for OpenCode as `~/.local/share/opencode/opencode.db` using `filepath.Join` with proper path handling. The database is opened in read-only mode (`mode=ro`) using a `file:` URI prefix to ensure the `modernc.org/sqlite` driver enforces the read-only constraint. Without the `file:` prefix, query parameters like `mode=ro` are silently ignored and the driver opens in read-write mode.
 

docs/DREAM.md

@@ -8,7 +8,7 @@ The dream cycle performs automated memory maintenance during idle periods. It ca
 
 - **Light phase:** Sort and deduplicate near-duplicate memories (cosine similarity ≥ threshold).
 - **Deep phase:** Score, promote, decay, and merge memories. Decays confidence on idle memories. Boosts frequently accessed memories. Auto-links memories with high cosine similarity (≥ `dream.auto_link_threshold`, default 0.85) by creating `related_to` relations between them. Procedure memories older than `dream.stale_procedure_days` (default 30 days) with no recent access decay at double the normal rate — proposed-but-never-adopted procedures fade faster than confirmed ones.
-- **REM phase:** Extract themes from memory clusters and write a dream diary (read-only reflection). Also extracts behavioral insights (patterns exceeding `dream.behavioral_threshold` occurrences within `dream.behavioral_lookback_days` days). When Ollama is available, uses an LLM call to generate specific, actionable procedural rules with "Do" directives and "Verify" steps. Falls back to count-based summaries when Ollama is unavailable.
+- **REM phase:** Extract themes from memory clusters and write a dream diary (read-only reflection). Also extracts behavioral insights (patterns exceeding `dream.behavioral_threshold` occurrences within `dream.behavioral_lookback_days` days). When Ollama is available, uses an LLM call to generate specific, actionable procedural rules with "Do" directives and "Verify" steps; also generates `[SKILL PATCH]` sections (Detection Rule, Checklist, Pitfall, Verification). Falls back to count-based summaries when Ollama is unavailable. When run with `--apply`, appends behavioral insight and skill patch sections to `proposed-changes.md`, with a timestamp header separating dream runs. Proposed procedure memories are linked to their `proposed-changes.md` entry via `proposed_changes_link` metadata.
 
 Configuration is under the `dream:` key in `config.yaml`. See [Configuration](CONFIGURATION.md) for all dream settings.
 
@@ -48,6 +48,7 @@ dreamer, err := dream.NewDreamer(dream.DreamerConfig{
     OllamaClient:          nil,    // nil → created from BaseURL; takes precedence if provided
     DiaryPath:             "",     // defaults from paths.GetDreamDiaryPath()
     ReportPath:            "",     // defaults from paths.GetDreamReportPath()
+    ProposedChangesPath:   "",     // defaults from paths.GetProposedChangesPath()
 })
 if err != nil {
     log.Fatal(err)
@@ -62,6 +63,9 @@ result, err := dreamer.Run(ctx, true, "deep")
 // Write dream diary (markdown with sync.Mutex for in-process concurrency)
 err = dreamer.WriteDiary(result)
 
+// Write proposed-changes.md (behavioral insights + skill patches, append-only)
+err = dreamer.WriteProposedChanges(ctx, result)
+
 // Generate HTML dream report
 err = dreamer.GenerateDreamReport(result, "/path/to/report.html")
 ```
@@ -88,6 +92,7 @@ err = dreamer.GenerateDreamReport(result, "/path/to/report.html")
 | Model | string | `"glm-5.1:cloud"` | Ollama model name for behavioral insight generation |
 | DiaryPath | string | paths.GetDreamDiaryPath() | Path for dream diary markdown |
 | ReportPath | string | paths.GetDreamReportPath() | Path for HTML dream report |
+| ProposedChangesPath | string | paths.GetProposedChangesPath() | Path for proposed-changes.md (behavioral insights and skill patches) |
 
 #### DreamResult
 
@@ -117,6 +122,7 @@ type BehavioralInsight struct {
     Count          int
     InsightID      string
     ContentSnippet string
+    Samples        []string
 }
 
 type RemPhaseResult struct {

internal/config/config.go

@@ -42,7 +42,8 @@ type DreamConfig struct {
 	Model                  string  `yaml:"model"`
 	// ModelTimeout is the timeout for each LLM call during REM behavioral insight generation.
 	// Parsed as a Go duration string (e.g. "5m", "120s"). Defaults to "5m".
-	ModelTimeout string `yaml:"model_timeout"`
+	ModelTimeout        string `yaml:"model_timeout"`
+	ProposedChangesPath string  `yaml:"proposed_changes_path"`
 }
 
 // SessionConfig holds session lifecycle settings.
@@ -239,6 +240,7 @@ func (c *Config) DreamerConfig() dream.DreamerConfig {
 		StaleProcedureDays:     c.Dream.StaleProcedureDays,
 		DiaryPath:              c.Dream.DiaryPath,
 		ReportPath:             c.Dream.ReportPath,
+		ProposedChangesPath:    c.Dream.ProposedChangesPath,
 		BaseURL:                ollamaURL,
 		Model:                  model,
 		ModelTimeout:           modelTimeout,