Memory

Five-layer agent memory.
Each client has its own context.

MemGPT-inspired architecture that the agent self-edits. No cold start, no blank-page hallucinations.

L1

Core Memory

always in prompt

Client, targets, strategy, guardrails. Loaded into every agent prompt.

L2

Episodic memory

vector-indexed

Every approved/rejected change, A/B test, anomaly and your note. Your "no" weighs 3× more than automated data.

L3

Patterns & ontology

Bayes · 146 concepts

Rules like "tCPA under 30 conversions raises CPA by 20–40%". Bayesian scores with confidence breakdown.

L4

Knowledge graph

Memgraph · multi-hop

Causal chains. Multi-hop queries through relationships, not just vectors.

L5

Autodream

after each action · Monday full

Continuous memory improvement + weekly consolidation. Stale patterns decay, confirmed ones strengthen.

Example: pause a seasonal keyword

A search term "winter passes 2024 ski" — tCPA rising, ROAS dropping. The agent finds a note in L2 from March 2025: "demand drops in April, we did not pause — pause to save budget". L3 confirms the pattern "seasonal down-trend April → May for winter terms, weight 87%". Proposal: pause, restart Dec 1. You approve, audit log saves a snapshot for rollback.

Deeper explainer

Detailed article on each layer, the ontology, Bayes and autodream — 12 minute read.