PTN—42
Relationship continuity
Carry context, preferences, and unfinished work across sessions so every return doesn’t start from zero.
Most AI products have amnesia by architecture: each session is a fresh context window, and the burden of continuity falls on the user, who re-explains the project, the constraints, and the taste they already taught it last Tuesday. Relationship continuity is the pattern of carrying that accumulated state forward deliberately — durable workspaces, resumable threads, standing instructions — so the hundredth session is meaningfully better than the first.
The workhorse structure is the project or workspace: a durable container that holds instructions, attached files, and conversation history for one ongoing effort. Claude’s Projects and ChatGPT’s equivalent both landed on this shape because it solves the re-explaining problem without global memory’s creepiness — context is scoped to the effort, visible in the container, and dies with the project instead of following you everywhere.
Continuity has a second half teams skip: re-entry. Returning to a three-week-old thread means returning to a decision state you’ve forgotten. A good re-entry surface summarizes where things stood — what was decided, what was produced, what was left open — instead of dumping you at message 147. Granola does this for meetings; agent products need it for work sessions. The transcript is the record; the summary is the doorway.
The discipline is knowing what not to carry. Preferences and project facts compound; moods, one-off tangents, and abandoned directions don’t. Continuity that replays everything becomes clutter, and clutter in an AI context window isn’t just noise — it actively degrades the answers. Curate what persists, and let people see and prune it, which is where this pattern hands off to legible memory.
When to use
- People return repeatedly for ongoing work — a codebase, a campaign, a manuscript.
- Sessions build on prior decisions that are expensive to re-establish by hand.
- Long gaps between sessions make “where were we?” a real cost on re-entry.
- Preferences taught once (tone, format, stack) should hold everywhere they apply.
Design considerations
- 01Scope continuity to projects or workspaces before reaching for global memory — scoped context is legible by construction.
- 02Greet returning users with a state summary: decisions made, artifacts produced, open threads.
- 03Let standing instructions live as editable text inside the container, not as hidden profile state.
- 04Make threads resumable and findable — search over past conversations is continuity infrastructure, not a nice-to-have.
- 05Expire or archive stale context deliberately; a six-month-old constraint silently applied is a bug.
- 06Keep continuity portable across surfaces — the preference set on desktop should hold on mobile and in the CLI.
Pitfalls
- ✕Continuity as clutter: every past tangent rides along in context, answers drift, and users open fresh chats to escape their own history — the product’s continuity feature is now the thing people route around.
- ✕Cross-project leakage — tone rules from a legal project bleeding into a birthday-poem chat. Scoping failures read as the system gossiping about you to yourself.
- ✕Resuming without recap: the thread reopens at the bottom of a wall of messages and the user must archaeologize their own decisions.
- ✕Treating continuity as retention mechanics — hoarding context to raise switching costs rather than to improve answers. Users can tell, and export requests are how they tell you.
In the wild
- Claude
- Projects bundle instructions, files, and chats into a durable container — context is scoped to the effort and readable in place, so continuity never depends on invisible state.
- ChatGPT
- Combines persistent memory, custom instructions, and searchable chat history so preferences and past work carry across sessions.
- Cursor
- Rules files pin project conventions as version-controlled text — standing instructions that persist across every session and every teammate.
- Granola
- Builds continuity across meetings: past notes inform new ones, and recaps give you the re-entry summary instead of a raw transcript.
- Linear
- Its agent workflows attach AI work to issues — the durable object is the issue itself, so context and history live where the team already looks.