LoomDB

The question

Can memory retrieval be modeled as spreading activation over a temporal graph — where forgetting is a feature, not a failure?

Overview

Memory as a nonlinear dissipative dynamical system over a mutable directed graph — not a passive vector store. Activation saturates asymptotically, stability grows sublinearly, and sub-stochastic edge weights guarantee bounded signal under arbitrarily long horizons: repeated stimulation and cycles cannot explode the state.

Approach

• 1.Asymptotic saturation: a_v(t+1) = a_v(t) + (1 - a_v(t))α — every stimulus contributes a fraction of the remaining distance to the ceiling, so energy never accumulates unbounded.
• 2.Lazy decay modulated by consolidated stability: frequently reinforced memories resist forgetting; neglected ones fade on a multiplicative schedule.
• 3.Sub-stochastic propagation: outgoing weights sum to ≤1, making boundedness a structural property of the graph, not a runtime hope.
• 4.Evolutionary policy optimization: the hybrid semantic+activation reranking policy is treated as an evolvable program — a small interpretable router over query features discovers when to trust text similarity vs. activation energy, instead of fixed magic numbers.
• 5.Rust core compiled to WASM — the same engine runs server-side, in the browser, and embedded in desktop agents.

Open problems

Stated plainly, because pretending they're solved would be the opposite of research.

• Consolidation: when should episodic traces merge into semantic structure?
• Decay calibration across radically different interaction frequencies.
• Benchmarking long-horizon recall without long-horizon datasets.