Sediment Palace: Local Memory for AI Agents Modeled on Geological Layers

In 2025, AI agents have reached impressive heights: they write code, analyze documents, negotiate and operate tools. But most of them share a fundamental weakness: memory. More precisely, they either have none between sessions or rely on extremely crude ways of organizing it.

A developer who published a write-up on Habr took two existing concepts and combined them into something new: a system called Sediment Palace. The core of the model is an analogy with geology. Memory is organized not as a flat list of records and not as a vector space, but as a layered structure in which data literally "settles" over time.

Fresh information, meaning events from the last minutes or hours, sits in the top layer. It is quickly accessible, takes up a lot of space and is uncompressed. As time passes, records sink deeper: they are compacted and aggregated, losing detail but keeping the essence. The oldest layers are like fossils: still accessible, but heavily compressed and consulted less often.

This is a direct analogy with how human memory works. We remember clearly what we ate for breakfast today, vaguely what happened a month ago, and keep only compressed "imprints" of events from years back. The brain does exactly what Sediment Palace tries to reproduce in software: it automatically compacts old experience while preserving its key patterns.

Unlike classic approaches, such as RAG systems on vector databases or a simple context window, the sedimentation model works locally and deterministically. It needs no external database API, no probabilistic search over embeddings, and does not depend on the size of the model's context window. The agent itself manages what it remembers and in what form.

The author admits the idea is not entirely original. He took two existing approaches, the concept of hierarchical memory and the metaphor of geological layers, and joined them into a working architecture. That candor is what makes the write-up valuable: the AI tooling space is full of work that reinvents the wheel without citing predecessors. Here, the opposite is true.

The system's practical value is most visible in scenarios with long-lived agents: personal assistants, automated researchers, monitoring agents. For tasks where a session lasts not minutes but days or weeks, Sediment Palace offers a compromise between completeness of history and the cost of storing it.

An important question arises when you look at the concept: how does the agent decide what to compact and what to discard as data moves between layers? This is the point where architectural elegance runs into practical difficulty. Any heuristic of "importance" is inevitably subjective, and this is the main challenge for systems like this.

Still, the very appearance of such projects is telling. The developer community is increasingly focused not on model power but on the infrastructure around models. Memory, tools and orchestration are the next frontier. Sediment Palace is one experiment on that frontier, and it deserves attention.