Longitudinal Memory

Even with full space referencing (which I understand is on the roadmap), the AI layer will remain session-scoped; it can read content, but it cannot accumulate understanding of it across interactions. Each session starts blank.

This is a fundamental limitation for any use case where the value of AI comes from knowing your history, not just your current content. The most useful thing an AI can do inside a PKMS is say: “this pattern has appeared in your writing, notes, or topics here’s when and how.” That requires memory that persists and grows, not just access to a large context window.

What’s needed is a user model that builds incrementally from content over time. Ability recognising recurring themes, behavioural patterns, emotional registers, and connecting current entries to relevant historical ones without the user having to direct the search manually.

This is architecturally different from full space referencing. Space referencing gives the AI access to read content on demand. Longitudinal memory means the AI develops and retains an evolving understanding of that content between sessions.

Without it, AI in Capacities remains a sophisticated lookup tool. With it, it becomes something genuinely closer to a thinking partner that knows you.

Comparable reference point: Claude’s Projects feature maintains persistent context across conversations through a combination of project knowledge files and conversation memory. The equivalent in Capacities would be an AI layer that indexes and retains pattern-level understanding of a user’s space over time, updated incrementally as new content is added.