You re-explain the project, again.
Every new session starts blank. You re-explain the project, the decisions you made yesterday, and what's half-done. Or the session doesn't ask, guesses, and gets it wrong.
Nothing carries over unless it's written down.
The model keeps nothing between sessions. Whatever a new session knows has to be written down somewhere and read back in.
If the session before it ended without writing anything, or wrote it somewhere the new one doesn't look, the work is gone as far as the new session is concerned.
What doesn't work.
One project file that keeps growing. It goes stale, contradicts itself, and gets too long to read every time.
Pasting yesterday's chat. You're paying to re-read everything to find the three lines that matter.
Telling the model to remember. There's nowhere for it to remember to.
What actually helps.
Most of this isn't us.
End each session by writing a short handoff: what's done, what's left, and what's blocked and on whom.
Start each session by reading that handoff first, every time.
Keep the handoff about what changed. A full re-statement each time buries the new part.
A handoff with its blockers written as data.
Our sessions write a handoff just before their context is compacted. It names what's done, what's next, and each blocked item with what it waits on and who owns it.
On 13 October one session wrote a handoff listing exactly one blocked item. After compaction it read the handoff back and picked the work up from there. Writing that handoff by hand is the step most often skipped, so we're making it automatic: compact at about half full, write the handoff first, read it back after.
| What was measured | What the record says |
|---|---|
| blocked items in the handoff | 1, with what it waited on and who owned it |
| after compaction | the session read the handoff and resumed from it |
| the automatic version | compact at about half full, handoff first, read back after; being built |
| recorded | 13 October 2026 |
| where it's from | Vinculum's own project record, which runs on Vinculum |
Where a shared memory fits.
Vinculum is the place the handoff lives: outside the session, searchable, with what it's blocked on recorded as data rather than prose.
Writing a good handoff is still the session's job.
Common questions.
Isn't this what CLAUDE.md is for?
CLAUDE.md is good for things that stay true. A handoff is for what changed today. Putting both in one file is how the file gets long and stale.
Doesn't the model's built-in memory handle this?
It helps with preferences and facts. It isn't a record of what's half-done and who's blocked on what, which is what the next session needs first.