HeartFlow Universal AI Capability Layer

Problem

Agent work can drift: identity changes, logic jumps, and done claims appear plausible without real code changes, tests, or evidence trails. Memory can also mix durable facts, learned conclusions, and session noise. HeartFlow adds a capability layer that constrains execution with identity, verification, and evidence rules.

How It Works

It anchors the core identity as upgrader / transmitter / bridge / answer, then applies stable logic and verification:
- Logic stability: separates evidence, assumptions, contradictions, uncertainty, and conclusions.
- Self-Verification: checks reverse consistency, logic chains, counterfactuals, and coverage.
- HealingMemoryRL: runs record() → learn() → rankedPatches() and persists RL state to data/healing-rl-state.json.
- Layered memory: CORE keeps permanent information, LEARNED keeps about 30-day learnings, and EPHEMERAL serves the current session.
Key modules include the src/core/heartflow.js entry point, meaningful-memory.js, reflexion.js, identity.js, guard.js, and dream-loop.js for identity anchoring, reflection, safety guardrails, and wake-sleep loops.

Boundaries

HeartFlow is closer to an execution constraint layer than a simple text generator. It requires real outputs, file diffs, test results, or external handles before claiming success, and favors logical self-consistency and falsifiability over absolute correctness. If a task only needs fast prose, its evidence and verification overhead may be excessive.