feat: EchoForge — Temporal Predictive Consolidation Engine with Causal Ripple Simulation

[Copilot]

TIMPs' reflection layer is reactive (post-hoc LLM scoring adds 400–900ms), lacks hierarchical memory abstraction for longitudinal fidelity, and has no mechanism for simulating future states (burnout trajectories, relationship drift, contradiction cascades). EchoForge addresses all three by building a temporal memory DAG on the existing PostgreSQL/Qdrant dual-search layer.

Core Engine (core/echoForge.ts)

• Write-time gating — scores incoming memories via importance × novelty × tag_impact proxy (no LLM call). Gates below threshold 0.3 to skip low-value ingestion
• DAG graph connection — connects new nodes to recent neighbors via temporal edges (30-day window decay) and semantic edges (tag overlap ≥ 0.3). O(k) bounded, k=10
• Causal ripple simulation — bounded propagation (depth 5) on the DAG using recursive CTEs. Detects burnout risk, relationship drift, contradiction cascades with probabilistic scoring. Stores predictions when risk ≥ 0.6
• Hierarchical consolidation — periodic batch merge of raw → episodic summary nodes by tag cluster, with Ebbinghaus-inspired temporal decay (0.995^t) and pruning of fully decayed nodes
• Proactive insights — pre-computed warning strings injected into agent system prompt context

Database Schema (db/echoForgeDb.ts)

Five new tables: echoforge_dag_nodes, echoforge_dag_edges (with unique constraint + self-loop prevention), echoforge_trajectories (auto-expiring via horizon_days), echoforge_ripple_results, echoforge_consolidation_log.

Tool 18 (tools/echoForgeTool.ts)

Operations: predictions, simulate, consolidate, trajectories, insights, status.

Integration

• agent.ts — memories ingested into DAG during reflectAndStore; proactive insights appended to context in run()
• toolRouter.ts — keyword routes for predict|forecast|trajectory|ripple|echoforge|anticipate|proactive
• toolsDb.ts — initEchoForgeTables() called during migration

// During reflection, every stored memory is also ingested into the temporal DAG
const stored = await memoryIndex.storeMemory(...);
await echoForge.ingest(this.userId, stored.id, memory.content, memory.tags, memory.importance);

// Proactive insights injected into agent context before each interaction
const insights = await echoForge.getProactiveInsights(this.userId);
// → "⚠️ EchoForge Alert [burnout_risk]: Risk 72% (confidence 65%) — 4 stress signals in 12 connected memories"

Zero new dependencies. All EchoForge paths are supplementary with silent failure handling — existing behavior unchanged if tables don't exist or queries fail.