A new kind of architecture

Agents have, until now, typically been stateless. This means they have no memory at either the beginning or the end of a given session; each interaction started from zero. However, there's been a shift away from that model and toward something persistent: systems that retain knowledge between sessions, consolidate experience in the background and maintain a layer of latent context that shapes behavior without being explicitly loaded into any single prompt. The analogy should be clear; the accumulated weight of prior experience which operates beneath the surface of active thought forms a kind of 'agent unconscious'.

The technical foundation of this shift became visible in March, when an accidental npm packaging error shipped 512,000 lines of Claude Code TypeScript source code. The analysis published by AlphaSignal revealed an architecture that hadn’t been publicly documented before. At its core there was a three-layer memory system built around a compact pointer file — a lightweight index that tells the agent what it knows and where to find it, without loading the content until a query demands it. The agent has access to knowledge but it doesn’t carry it around.

Tackling context pollution

This provides a solution to context pollution. As we pointed out in the recent Technology Radar, urging caution around ‘agent context bloat’, filling a context window with everything that might be relevant degrades reasoning rather than improving it. The pointer-and-retrieval architecture, though, keeps active context lean while giving the agent access to a knowledge base far larger than any single prompt could hold. Domain knowledge lives in indexed topic files — a skills layer —which is only injected when it’s relevant. What remains outside the active window is latent; it’s available when needed but doesn’t occupy the agent’s attention.

The implications for organizations

The organizational implication is even more significant than the technical one because the agent's unconscious needs to be built; it doesn’t, of course, emerge organically. Someone has to decide what enters the index, how domain knowledge is structured, which pointers are maintained and how the skills layer is curated over time. This isn’t primarily a software problem; it’s an organizational knowledge architecture problem which is now becoming a prerequisite for any agent deployment where institutional judgment matters.