Markov Prediction#

Parametric Memory includes a Prediction by Partial Matching (PPM) Markov model trained on your own recall sequences. It observes which atoms you access and in what order, then pre-fetches atoms it predicts you'll need next — before you ask for them.

How it works#

When you recall atoms, the sequence is recorded as a training sample:

session: recall(A) → recall(B) → recall(C) → recall(A)

After session_checkpoint, the model trains on this sequence. Future recalls of A will prime a prediction that B is likely next.

Prediction in practice#

// After recalling one atom, ask for predictions
const recall = await memory.recall("v1.user.Glen.preference.theme");
 
// Prefetch predicted-next atoms
const predictions = await memory.predict({ after: recall.atom.key, limit: 3 });
// → ['v1.user.Glen.preference.language', 'v1.user.Glen.tools.editor', ...]

You can request predictions proactively at session start to warm your local cache.

Performance numbers#

Controlling the model#

session_checkpoint triggers a retrain. Call it at natural session boundaries — not on every store, as retraining is synchronous and briefly pauses writes.

// At end of a work session
await memory.sessionCheckpoint();
// → flushes log → trains → commits → updates Markov weights

When prediction helps most#

Prediction is most effective when:

• Keys use consistent namespaces (co-occurrence patterns are learnable)
• Sessions are long enough to generate training sequences (>5 recalls)
• You call session_checkpoint regularly so the model stays current