Skip to main content
Every memory diff carries an energy budget. That budget is the single dial that decides how far the diff spreads across the mesh during gossip propagation: the diff forwards to a neighbour only while it can still afford the hop, and every hop spends energy. This page gives you the exact formulas and the intuition behind them.

​
Initial energy

When a diff is created, its starting energy is:
  • base_energy is the global baseline (default 10.0).
  • salience is how important the diff is, in [0, 1]. The square root softens the effect, so a very low-salience diff still gets a meaningful budget rather than collapsing to zero.
  • type_multiplier depends on the cell type (see the table below). An unknown or unrecognised cell type uses a multiplier of 1.0.
A contradiction diff is special: its initial energy is multiplied by contradiction_bonus = 1.5. Conflicting information is given extra reach so it can catch up with whatever it contradicts.

​
Hop cost

Each hop spends energy based on the strength of the edge being crossed:
Stronger edges (higher weight) are cheaper to cross, so diffs naturally flow along the well-connected paths of the mesh. The edge weight is floored at 0.01 so that a near-zero weight cannot produce an infinite or undefined cost. A diff keeps forwarding across an edge only while diff.energy >= hop_cost. Once it can no longer afford any remaining edge, it stops.

​
Cell-type multipliers

The cell type sets how much initial energy a diff starts with — which is what makes some kinds of knowledge travel across the whole mesh while others stay local. The intuition is deliberate. Decisions and constraints shape everyone’s work, so they start with the most energy and reach the farthest. A task is local and short-lived, so it starts with little energy and stays near its origin. Facts sit in the middle at the neutral 1.0 baseline.
These multipliers are calibrated values, not arbitrary defaults. Changing them changes how far each kind of knowledge propagates across the mesh. Treat them as a tuned part of the system’s behaviour, not a casual configuration knob.

​
Worked intuition

Consider two diffs stored with the same salience across the same neighbourhood:
  • A Decision starts at 10.0 × √salience × 2.0 — a large budget. It can afford many hops and fans out across the mesh.
  • A Task starts at 10.0 × √salience × 0.3 — a small budget. It pays for a hop or two and then stops, staying close to where it was created.
That single difference in starting energy is why “decisions travel furthest” while routine tasks stay local — without any explicit routing rules.

​
How this connects to the rest of the system

Energy is the budget that bounds the gossip propagation walk. Trust can reduce it further: with trust.energy_attenuation enabled (default false — a stock deployment does not attenuate), a low-trust source’s effective energy is scaled down by its trust score. And layer promotion adds a fixed energy bonus when a diff graduates, giving the promoted diff fresh reach.

​
Source

This page is a teaching restatement of the HyphaeDB specifications. It does not define new behaviour.