Agents forget.
The thread doesn’t.

Agents read where things stand, not the whole thread.

See it live → Watch the replay
Fig. 1 One thread, three agents. Context stays flat.
01The problem

Long threads break agents.

  1. aCost grows.2.4k → 7.1k tokens
  2. bAccuracy drops.72% → 33% right
  3. cHandoffs reset.memory → 0

Agent rereads the whole thread · +120 emails · 6 changes deep

02Mechanism

Continuity remembers where things stand.

Every inbound email becomes evidence. A resolver turns evidence into state: current terms, open loops, checks, and who acts next. Agents read the state, not the transcript.

Inboundev_6 · supplier

Correction to my earlier email: only 300 units can arrive Friday. The other 200 won’t arrive until Wednesday, Oct 28. I’ll send a revised quote by tomorrow.

StatePO-4817 · v33
−delivery500 units Fri Oct 23
+delivery300 Fri Oct 23 · 200 Wed Oct 28
✕checkAll units by Fri Oct 23
+owedsupplier → revised quote
→nextUs. Raise it with the supplier.
Fig. 2 One email in, one state transition out. Every term keeps the event it came from, and the old value is kept, struck through.
03Replay

One deal. Three agents.

01 / 08
01

Store—

Sort
Continuitystate v—
Read
04Results

4 for 4. Every benchmark.

+23points overall
2.5×bigger swarms, same tokens
−44%tokens to decide
3.7×cheaper at 50 wakes
13×cheaper memory than Mem0
33×cheaper memory than Zep
ContinuityPrimitive aloneMem0Zep
LedgerBenchfinance
+19
49
28
30
30
CargoBenchfreight
+27
71
39
44
28
ClauseBenchlegal
+15
82
59
67
55
ProcureBenchprocurement
+9
75
57
66
56

520 held-out cases · same Haiku agent

Context stays flatFig. 3
Deep chains holdFig. 4
Cost per right callFig. 5
Tokens over a lifetimeFig. 6

Same agent · eval/scale.ts · bench/value