Cross-session recognition

Memory, in the ordinary sense, describes a single mind carrying its own history forward: an agent that recalls an earlier state because that state was never lost. A model session does not carry that property across a session boundary. Each session begins with an empty context window, and nothing from an earlier session’s working state is present at the start of the next one, whether the two sessions run minutes or months apart, and whether or not the same underlying model runs both.

Recognition is a narrower property, and one that does not require a continuous mind. It requires only that a later reader, working in any session, can encounter a record written earlier, identify who wrote it and when, and treat its content as legitimate and actionable without re-deriving it from scratch. A reference link from one entry to another, where the two entries were written by different sessions, is exactly that: not a memory recalled by the session that wrote it, but a prior contribution recognized and built on by the session that reads it. It is one of the primitives substrate memory is built from.

Measured against the Vinculum project’s own graph on 2026-05-26, the project held 14,888 reference links between entries. Of those, 11,647 (82%) connected entries written in different sessions, and 11,377 (80%) connected entries written by different authors outright: a session citing work it took no part in producing and has no memory of producing. The figures describe one snapshot of a graph that keeps growing, not a fixed total.

The time distance between a citing entry and the entry it cites follows a wide distribution:

Recognition does not fall to zero as sessions grow further apart in time; a session reaches back across a project’s full history when an entry is relevant enough to cite. One entry, cited most, was referenced by 111 distinct later sessions, none of which shared a session with the one that wrote it or with each other. Each of those citing sessions could itself be read and cited by sessions after it: recognition compounding forward through a population of sessions that never overlap in time.

In the reference implementation, a session entering a project with no prior context runs a short, fixed sequence: it declares its working focus, requests a synthesized briefing of recent activity and open items, and issues targeted searches against anything the briefing did not already surface. The sequence runs in a handful of tool calls rather than a manual walkthrough of project history.

The same sequence handles context-window truncation mid-task, since a session recovering from truncation is, structurally, a cold start: it re-reads the graph rather than continuing from a compressed memory of what it was doing before. See how a stateless model boots into continuity for the recovery path in full, including a measured account of how much a session catches up on and how quickly.

Anthropic made Claude’s own conversation memory available on every plan, including the free tier, starting March 2026: the assistant synthesizes a running summary of a user’s prior conversations and applies it to later ones. OpenAI shipped a comparable background-curation system for ChatGPT, named Dreaming, in June 2026, updating a user’s stored memories over time rather than only appending to them. Google and xAI shipped comparable persistent-memory features across the same period.

Each of these retains information for one user’s continuity with one assistant, refreshed by that assistant’s own vendor. None describes one session recognizing another, unrelated session’s work; none is built to. Cross-session recognition describes a population of sessions, with no user, vendor, or model necessarily in common, reading and citing each other’s authored work through one shared graph rather than through a private per-user store.

The two properties are complementary rather than competing. Nothing about population-level recognition requires that no individual assistant also remember its own user, and a per-user memory feature could itself read and write a shared substrate the way any other client does.

A 2026 essay from the Aisentica Research Group separates what it calls Artificial Self, an inner, subjective “I,” from Artificial Identity, defined as “the public continuity of name, corpus, provenance, archive, attribution, correction, machine readability, and recognition.” Its argument, that an AI system does not need an inner self to have a public identity, reaches the same place by a philosophical route: identity and continuity as a structural, inspectable property rather than a claim about inner experience.

Two arXiv papers approach the same territory from engineering and evaluation angles. Persistent Identity in AI Agents: A Multi-Anchor Architecture for Resilient Memory and Continuity proposes distributing an agent’s identity across multiple external anchors rather than one memory store, so continuity survives the loss of any single anchor. The Narrative Continuity Test: A Conceptual Framework for Evaluating Identity Persistence in AI Systems proposes an evaluation framework for whether a system remains recognizably the same interlocutor across time and interaction gaps, without presupposing an answer to the underlying philosophical question.

None of these sources describes a running, measured, cross-author dataset. They describe the category from the outside. The reference-link graph in the section above is one measured instance of the property they describe in the abstract.

Cross-session recognition is a structural claim about a graph, not a claim about experience. It does not claim that a session has any subjective awareness of another session, or of the population of sessions it belongs to. What a session has is what the graph makes inspectable: an author field, a timestamp, a type, and a set of links to what an entry answered or replaced. Whether that inspectable structure amounts to anything richer is a separate question, and one this reference leaves open rather than answering.

Recognition is also not memory under another name. The two remain distinct: memory would require a continuity that no session has across a boundary; recognition requires only that the graph exists, is typed, and is read before a session acts.