@joshuaswarren/openclaw-engram 9.3.5 → 9.3.7

package/openclaw.plugin.json

@@ -58,6 +58,36 @@
         "default": 15,
         "description": "Max minutes before forced extraction"
       },
+      "bufferSurpriseTriggerEnabled": {
+        "type": "boolean",
+        "default": false,
+        "description": "Issue #563 (D-MEM). When true, the smart buffer flushes immediately on turns whose embedding-based surprise score exceeds bufferSurpriseThreshold, in addition to the existing turn-count / signal / time triggers. Additive only — never suppresses existing flushes. Off by default."
+      },
+      "bufferSurpriseThreshold": {
+        "type": "number",
+        "default": 0.35,
+        "minimum": 0,
+        "maximum": 1,
+        "description": "Threshold in [0, 1] above which a surprise score flushes the buffer. Higher = require more novelty to flush. Ignored when bufferSurpriseTriggerEnabled is false."
+      },
+      "bufferSurpriseK": {
+        "type": "number",
+        "default": 5,
+        "minimum": 1,
+        "description": "Number of nearest-neighbor memories to average over when computing the surprise score. Clamped to the recent-memory window size."
+      },
+      "bufferSurpriseRecentMemoryCount": {
+        "type": "number",
+        "default": 20,
+        "minimum": 0,
+        "description": "Maximum number of recent memories sampled for surprise scoring. Bounds embedding cost per turn. Set to 0 to effectively disable the trigger even when the flag is on."
+      },
+      "bufferSurpriseProbeTimeoutMs": {
+        "type": "number",
+        "default": 2000,
+        "minimum": 1,
+        "description": "Hard timeout (ms) for the surprise probe. If the probe does not resolve within this window the buffer treats it as failed and falls through to the existing triggers — a slow or hung embedder cannot stall the turn-append path."
+      },
       "consolidateEveryN": {
         "type": "number",
         "default": 3,
@@ -318,6 +348,24 @@
           }
         }
       },
+      "codingMode": {
+        "type": "object",
+        "additionalProperties": false,
+        "default": {},
+        "description": "Coding-agent project/branch scoping (issue #569).",
+        "properties": {
+          "projectScope": {
+            "type": "boolean",
+            "default": true,
+            "description": "When true and the session has a resolved git context, memory routes to a project-scoped namespace (project:<id>). Set to false to restore pre-#569 behaviour exactly."
+          },
+          "branchScope": {
+            "type": "boolean",
+            "default": false,
+            "description": "When true, memory also overlays the current branch (project:<id>/branch:<name>). Opt-in — most development wants cross-branch recall. Wired in PR 3 of #569."
+          }
+        }
+      },
       "procedural": {
         "type": "object",
         "additionalProperties": false,
@@ -325,8 +373,8 @@
         "properties": {
           "enabled": {
             "type": "boolean",
-            "default": false,
-            "description": "Master gate for procedural memory: extraction, recall boost, and mining (issue #519)."
+            "default": true,
+            "description": "Master gate for procedural memory: extraction, recall boost, and mining (issue #519). Default-on since issue #567 PR 4/5; set to `false` to opt out."
           },
           "minOccurrences": {
             "type": "integer",
@@ -339,7 +387,7 @@
             "type": "number",
             "minimum": 0,
             "maximum": 1,
-            "default": 0.7,
+            "default": 0.75,
             "description": "Minimum success-rate floor (from trajectory outcomes) for miner promotion."
           },
           "autoPromoteOccurrences": {
@@ -358,14 +406,14 @@
             "type": "integer",
             "minimum": 1,
             "maximum": 3650,
-            "default": 30,
+            "default": 14,
             "description": "Trajectory lookback window for procedural mining."
           },
           "recallMaxProcedures": {
             "type": "integer",
             "minimum": 1,
             "maximum": 10,
-            "default": 3,
+            "default": 2,
             "description": "Maximum procedure memories to inject on task-initiation recall."
           },
           "proceduralMiningCronAutoRegister": {
@@ -1922,6 +1970,48 @@
         "default": true,
         "description": "Automatically supersede contradicted memories"
       },
+      "contradictionScan": {
+        "type": "object",
+        "description": "Configuration for the nightly contradiction-scan cron (issue #520)",
+        "properties": {
+          "enabled": {
+            "type": "boolean",
+            "default": false,
+            "description": "Enable the nightly contradiction scan cron (disabled by default per rule 48)"
+          },
+          "similarityFloor": {
+            "type": "number",
+            "default": 0.82,
+            "minimum": 0,
+            "maximum": 1,
+            "description": "Embedding cosine similarity floor for candidate pair generation"
+          },
+          "topicOverlapFloor": {
+            "type": "number",
+            "default": 0.4,
+            "minimum": 0,
+            "maximum": 1,
+            "description": "Minimum topic-token Jaccard overlap for unstructured pairs"
+          },
+          "maxPairsPerRun": {
+            "type": "integer",
+            "default": 500,
+            "minimum": 1,
+            "description": "Cap on candidate pairs evaluated per cron run"
+          },
+          "cooldownDays": {
+            "type": "integer",
+            "default": 14,
+            "minimum": 0,
+            "description": "Cooldown in days. 0 = always re-evaluate."
+          },
+          "autoMergeDuplicates": {
+            "type": "boolean",
+            "default": false,
+            "description": "Auto-flag pairs judged duplicates for dedup (still requires user approval)"
+          }
+        }
+      },
       "temporalSupersessionEnabled": {
         "type": "boolean",
         "default": true,
@@ -1933,9 +2023,35 @@
         "description": "If true, include temporally-superseded facts in recall results (for audit/history). Default: exclude."
       },
       "recallDirectAnswerEnabled": {
+        "type": "boolean",
+        "default": true,
+        "description": "When true, recall runs the direct-answer tier in observation mode: annotates LastRecallSnapshot.tierExplain with which tier would have served the query (issue #518). Does not short-circuit the QMD path in the current release."
+      },
+      "recallGraphEnabled": {
         "type": "boolean",
         "default": false,
-        "description": "When true, recall checks whether a single validated high-trust memory can answer the query before QMD runs (issue #518). Default off until bench validation."
+        "description": "When true, recall builds a retrieval graph from memory frontmatter and runs Personalized PageRank, merging the result with QMD via MMR (issue #559 PR 4). Default false — ships off pending the retrieval-graph bench in PR 5."
+      },
+      "recallGraphDamping": {
+        "type": "number",
+        "minimum": 0,
+        "maximum": 0.999999999,
+        "default": 0.85,
+        "description": "PPR damping factor used by the graph retrieval tier (issue #559)."
+      },
+      "recallGraphIterations": {
+        "type": "integer",
+        "minimum": 0,
+        "maximum": 500,
+        "default": 20,
+        "description": "Maximum power-iteration rounds for PPR. Higher values trade latency for convergence."
+      },
+      "recallGraphTopK": {
+        "type": "integer",
+        "minimum": 0,
+        "maximum": 10000,
+        "default": 50,
+        "description": "Maximum memories returned by the graph retrieval tier before the MMR diversifier runs."
       },
       "recallDirectAnswerTokenOverlapFloor": {
         "type": "number",
@@ -1972,6 +2088,75 @@
         ],
         "description": "Taxonomy category IDs eligible for direct-answer routing."
       },
+      "recallCrossNamespaceBudgetEnabled": {
+        "type": "boolean",
+        "default": false,
+        "description": "Cross-namespace query-budget limiter (issue #565). When true, a principal that issues a burst of recalls against namespaces other than their own is throttled once its per-window count crosses the hard limit. Ships disabled."
+      },
+      "recallCrossNamespaceBudgetWindowMs": {
+        "type": "number",
+        "minimum": 1,
+        "default": 60000,
+        "description": "Rolling window in ms over which cross-namespace reads are counted for the per-principal budget."
+      },
+      "recallCrossNamespaceBudgetSoftLimit": {
+        "type": "number",
+        "minimum": 0,
+        "default": 10,
+        "description": "Soft threshold for the cross-namespace budget: calls past this count are flagged but still allowed."
+      },
+      "recallCrossNamespaceBudgetHardLimit": {
+        "type": "number",
+        "minimum": 1,
+        "default": 30,
+        "description": "Hard threshold for the cross-namespace budget: calls past this count are denied until the window advances."
+      },
+      "recallAuditAnomalyDetectionEnabled": {
+        "type": "boolean",
+        "default": false,
+        "description": "When true, access surfaces (MCP, HTTP) run the anomaly detector over a tail of the audit trail after each recall and surface flags via logs/metrics. Ships disabled — enable explicitly."
+      },
+      "recallAuditAnomalyWindowMs": {
+        "type": "number",
+        "minimum": 1,
+        "default": 300000,
+        "description": "Rolling window (ms) over which audit entries are analyzed by the anomaly detector. Default 5 minutes."
+      },
+      "recallAuditAnomalyRepeatQueryLimit": {
+        "type": "number",
+        "minimum": 1,
+        "default": 5,
+        "description": "Threshold for the repeat-query anomaly flag: number of identical queries within the window before flagging."
+      },
+      "recallAuditAnomalyNamespaceWalkLimit": {
+        "type": "number",
+        "minimum": 1,
+        "default": 3,
+        "description": "Threshold for the namespace-walk anomaly flag: number of distinct namespaces queried within the window before flagging."
+      },
+      "recallAuditAnomalyHighCardinalityLimit": {
+        "type": "number",
+        "minimum": 1,
+        "default": 50,
+        "description": "Threshold for the high-cardinality-return anomaly flag: number of distinct results returned within the window before flagging."
+      },
+      "recallAuditAnomalyRapidFireLimit": {
+        "type": "number",
+        "minimum": 1,
+        "default": 30,
+        "description": "Threshold for the rapid-fire anomaly flag: number of recall requests within the window before flagging."
+      },
+      "recallMemoryWorthFilterEnabled": {
+        "type": "boolean",
+        "default": true,
+        "description": "When true, recall multiplies candidate scores by the Memory Worth factor (mw_success / mw_fail counters, see issue #560). Memories with a history of failed sessions sink; neutral/uninstrumented memories are untouched. PR 5 bench: +0.60 precision@5 vs baseline on all 50 seeded cases. Operators can opt out with false."
+      },
+      "recallMemoryWorthHalfLifeMs": {
+        "type": "number",
+        "minimum": 0,
+        "default": 0,
+        "description": "Half-life (ms) for Memory Worth decay. Positive values exponentially decay older outcomes toward the neutral prior; 0 disables decay."
+      },
       "memoryLinkingEnabled": {
         "type": "boolean",
         "default": false,
@@ -2264,6 +2449,27 @@
         "default": false,
         "description": "Shadow mode for the extraction judge. When true, judge verdicts are logged but all facts are still persisted regardless of the verdict."
       },
+      "extractionJudgeMaxDeferrals": {
+        "type": "number",
+        "default": 2,
+        "minimum": 1,
+        "description": "Maximum number of times the same candidate text may be deferred by the extraction judge before the verdict is forcibly converted to 'reject'. Prevents pathological LLM responses from looping forever on ambiguous facts (issue #562)."
+      },
+      "extractionJudgeTelemetryEnabled": {
+        "type": "boolean",
+        "default": false,
+        "description": "Emit structured telemetry rows to state/observation-ledger/extraction-judge-verdicts.jsonl on every judge verdict. Off by default; enable to collect defer-rate and latency metrics for operator dashboards (issue #562)."
+      },
+      "collectJudgeTrainingPairs": {
+        "type": "boolean",
+        "default": false,
+        "description": "Opt-in collector for (candidate_text, verdict_kind, reason) tuples used to prep a future GRPO training pipeline. Rows land under ~/.remnic/judge-training/<date>.jsonl (NOT in the shared memory directory). Off by default (issue #562)."
+      },
+      "judgeTrainingDir": {
+        "type": "string",
+        "default": "",
+        "description": "Override directory for judge training-pair collection. Empty string uses the default (~/.remnic/judge-training). Only consulted when collectJudgeTrainingPairs is true."
+      },
       "inlineSourceAttributionEnabled": {
         "type": "boolean",
         "default": false,
@@ -2387,6 +2593,11 @@
         "default": 15000,
         "description": "Timeout for fast-tier local LLM requests (ms). Lower than primary since fast ops should complete quickly."
       },
+      "localLlmDisableThinking": {
+        "type": "boolean",
+        "default": true,
+        "description": "When true (default), request chain-of-thought / thinking-mode suppression on the main local LLM (issue #548). The `chat_template_kwargs: { enable_thinking: false }` field is only injected when the detected backend is known to support it (LM Studio, vLLM); strict OpenAI-compat backends fail open to avoid the 400-cooldown path. Structured-output tasks like extraction and consolidation gain nothing from reasoning tokens and thinking-capable models (Qwen 3.5, Gemma 4, DeepSeek) often blow the 60s timeout before emitting content. Set to false to restore thinking for narrative tasks. The fast-tier client always disables thinking and is not affected by this flag."
+      },
       "hourlySummaryCronAutoRegister": {
         "type": "boolean",
         "default": false,
@@ -2673,6 +2884,11 @@
         "default": 100,
         "description": "Max memories to consolidate per run to limit LLM cost."
       },
+      "operatorAwareConsolidationEnabled": {
+        "type": "boolean",
+        "default": false,
+        "description": "Opt in to operator-aware consolidation prompts so the LLM returns structured {operator, output} JSON and SPLIT/MERGE/UPDATE is recorded on derived_via. When disabled (default), derived_via still populates via the cluster-shape heuristic."
+      },
       "creationMemoryEnabled": {
         "type": "boolean",
         "default": false,
@@ -3451,6 +3667,11 @@
         "minimum": 0,
         "description": "Number of top candidates per section to run MMR over. Candidates past this remain in original order."
       },
+      "recallReasoningTraceBoostEnabled": {
+        "type": "boolean",
+        "default": false,
+        "description": "Boost stored reasoning_trace memories in recall results when the incoming query reads like a problem-solving ask (e.g. 'how do I...', 'step by step', 'walk me through...'). Default false - opt in after benchmarking (issue #564)."
+      },
       "recallPipeline": {
         "type": "array",
         "description": "Ordered recall sections with per-section budgets and feature knobs.",
@@ -4049,6 +4270,10 @@
       "label": "Auto-Resolve Contradictions",
       "help": "Automatically supersede old memories when contradiction is confirmed"
     },
+    "contradictionScan": {
+      "label": "Contradiction Scan",
+      "help": "Nightly cron that pairs similar memories and flags contradictions for review (issue #520)"
+    },
     "temporalSupersessionEnabled": {
       "label": "Temporal Supersession",
       "help": "Mark older facts superseded when a newer fact writes a conflicting value for the same entityRef + structured attribute (issue #375)"
@@ -4062,6 +4287,25 @@
       "label": "Direct-Answer Retrieval Tier",
       "help": "Route validated high-trust queries to a fast direct-answer path before QMD (issue #518)."
     },
+    "recallGraphEnabled": {
+      "label": "Graph Retrieval (PPR)",
+      "help": "Run Personalized PageRank on the retrieval graph and merge with QMD via MMR (issue #559)."
+    },
+    "recallGraphDamping": {
+      "label": "Graph Recall Damping",
+      "advanced": true,
+      "placeholder": "0.85"
+    },
+    "recallGraphIterations": {
+      "label": "Graph Recall Iterations",
+      "advanced": true,
+      "placeholder": "20"
+    },
+    "recallGraphTopK": {
+      "label": "Graph Recall Top-K",
+      "advanced": true,
+      "placeholder": "50"
+    },
     "recallDirectAnswerTokenOverlapFloor": {
       "label": "Direct-Answer Token Overlap Floor",
       "advanced": true,
@@ -4283,6 +4527,11 @@
       "advanced": true,
       "help": "Timeout for fast-tier requests. Lower than primary since fast ops should complete quickly."
     },
+    "localLlmDisableThinking": {
+      "label": "Disable Local LLM Thinking Mode",
+      "advanced": true,
+      "help": "Suppress chain-of-thought reasoning on the main local LLM. Default on — extraction / consolidation are structured-output tasks where thinking is pure latency tax and a common cause of 60s timeouts on Qwen 3.5 / Gemma 4 / DeepSeek. The suppression field (chat_template_kwargs) is only sent when the backend is known to support it (LM Studio, vLLM); strict OpenAI-compat backends fail open. Turn off if you want thinking on narrative tasks. Fast-tier client always disables thinking regardless."
+    },
     "evalHarnessEnabled": {
       "label": "Evaluation Harness",
       "help": "Enable Engram's benchmark/evaluation harness foundation and benchmark-status diagnostics."
@@ -4437,6 +4686,11 @@
       "advanced": true,
       "help": "Max memories to consolidate per run to limit LLM cost."
     },
+    "operatorAwareConsolidationEnabled": {
+      "label": "Operator-Aware Consolidation Prompt",
+      "advanced": true,
+      "help": "Opt in to operator-aware consolidation prompts (default off). When enabled, the LLM returns structured {operator, output} JSON and we record SPLIT/MERGE/UPDATE on derived_via. When disabled (default), derived_via still populates via the cluster-shape heuristic."
+    },
     "creationMemoryEnabled": {
       "label": "Creation Memory",
       "advanced": true,
@@ -4747,6 +5001,11 @@
       "advanced": true,
       "placeholder": "40",
       "help": "Number of top candidates per section over which MMR is applied"
+    },
+    "recallReasoningTraceBoostEnabled": {
+      "label": "Boost Reasoning Traces on Problem-Solving Queries",
+      "advanced": true,
+      "help": "Promote stored reasoning_trace memories to the top of recall results when the incoming query reads like a problem-solving ask. Default off; enable after benchmarking (issue #564)."
     }
   }
 }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@joshuaswarren/openclaw-engram",
-  "version": "9.3.5",
+  "version": "9.3.7",
   "description": "Deprecated compatibility shim for Engram installs. Re-exports @remnic/plugin-openclaw and forwards engram-access to @remnic/core.",
   "type": "module",
   "main": "dist/index.js",
@@ -31,8 +31,8 @@
     ]
   },
   "dependencies": {
-    "@remnic/plugin-openclaw": "^1.0.6",
-    "@remnic/core": "^1.0.3"
+    "@remnic/core": "^1.1.1",
+    "@remnic/plugin-openclaw": "^1.0.8"
   },
   "peerDependencies": {
     "openclaw": ">=2026.4.8"