UnisonContext rot
Context rot is the quality degradation that hits AI agents as long, unfiltered context accumulates - and why externalizing memory beats stretching the window.
Context rot is the performance degradation that happens when AI agents process increasingly long, unfiltered context - and it's why long agentic sessions contradict their own earlier decisions.
The mechanism
Models don't attend uniformly across a long window. As a session accumulates tool output, dead ends, and stale plans, the signal the model needs competes with noise it can't discard. Quality drops before the window is technically full - more tokens, worse answers.
The fix is selection, not size
The robust pattern is to keep the working context short and externalize durable knowledge: write decisions and facts out as they happen, then re-inject only the relevant slice when needed. That is exactly the Unison loop - ingest as you go, recall at task boundaries, with weakEvidence abstention instead of padding.
A bigger window doesn't fix rot; it postpones it while raising cost. See Why a bigger context window doesn't solve agent memory.
Shared vs private memory
The Unison visibility model: write location decides who sees it - /private/ for the caller, /workspace/teams/ for a squad, /workspace/ for the whole workspace, with explicit promotion.
Team vs personal memory
Personal memory makes one agent remember one user. Team memory makes every agent and human on a project share one accumulating source of truth.