Bayesian Memory — Docs

TrueZone uses Bayesian accumulation to refine parameter estimates across multiple sessions. Each session produces a likelihood surface over (E, Vmax, P) space, which is combined with the prior from previous sessions to produce an updated posterior.

How it works

First session: The SDK grid-searches over E, Vmax, P to find the best fit. The error surface is stored as the memory.
Subsequent sessions: Pass the previous memory as input. The SDK combines the new session's likelihood with the prior, weighted by confidence and time decay.
Convergence: After 3–10 sessions, parameters stabilize and confidence reaches a plateau.

Memory format

{
  "secsSinceEpoch": 1712000000,
  "confidence": 3.5,
  "minKey": "18504265",
  "minValue": 12.5,
  "cutoffValue": 15.0,
  "keys": [18504265, 18504266, 18604265, ...],
  "errorsX1000": [12500, 13200, 14100, ...]
}

Field	Description
`secsSinceEpoch`	Timestamp of the last analysis (for time decay)
`confidence`	Accumulated confidence (0–10). Higher = more certain
`minKey`	Key of the best parameter combination
`minValue`	Lowest error in the surface
`cutoffValue`	Error threshold for pruning
`keys`	Array of parameter combination keys (integer-encoded)
`errorsX1000`	Corresponding errors, multiplied by 1000 and stored as integers

Key encoding

Each key encodes a (P, speed, E) triplet as a single integer. The speed component is an internal parameter — Vmax is derived from it and E.

key = P × 100000 + speed × 36 + E × 100

Confidence decay

Confidence decays over time to allow the model to adapt as fitness changes. The decay rate is set so that confidence drops by approximately 50% over one year of inactivity. Regular training maintains high confidence; extended breaks gradually allow the model to re-learn.

Bayesian update

posterior = (likelihood_confidence × likelihood + prior_confidence × decay × prior)
          / (likelihood_confidence + prior_confidence × decay)

The updated confidence is capped at 10.0. The memory retains up to 1000 parameter combinations, pruning low-probability regions.

Usage

Pass memory between sessions:

# Session 1
payload1 = {
    "settings": json.dumps({"fixedSpeed": False, "fit": True}),
    "data": json.dumps(session1_data)
}
result1 = requests.post(ENDPOINT, json=payload1,
                        headers={"Authorization": f"Bearer {TOKEN}"}).json()
memory = result1["memory"]

# Session 2 — pass the memory from session 1
payload2 = {
    "settings": json.dumps({"fixedSpeed": False, "fit": True}),
    "data": json.dumps(session2_data),
    "memory": json.dumps(memory)
}
result2 = requests.post(ENDPOINT, json=payload2,
                        headers={"Authorization": f"Bearer {TOKEN}"}).json()

# Use the Bayesian-refined estimates
true_e = result2["result"]["trueE"]
true_vmax = result2["result"]["trueVmax"]
true_p = result2["result"]["trueP"]

The trueE, trueVmax, trueP fields are the posterior (accumulated) estimates. Always use the true values for display and downstream calculations.