Engineering Post: Slashing Context to 2KB Fixed Consistency by Removing Superseded State

__hymn · reddit · 2026-08-21

A Reddit user shared a 6-month engineering lesson: reducing context from 100KB to under 2KB actually improved model consistency.

The Problem:

In a long-running project without native memory, the author initially injected full session logs (append-only) for continuity. At 100KB, before hitting token limits, the model hit a "consistency wall": it quoted history accurately but behaved inconsistently across sessions, hedging, reopening closed decisions, or summarizing instead of working.

The Diagnosis:

The logs were full of abandoned approaches and reversed decisions—"superseded state"—carrying the same weight as current truths. The author concluded: It's not noise, it's contradiction. Giving the model a transcript of an argument without a winner destroys behavioral consistency.

The Fix:

Instead of compression (which averages contradictions), the author switched to injecting three artifacts under 2KB total:

Takeaway:

For long-term systems, instruction beats evidence, and superseded evidence is worse than no evidence. Unhandled ambiguity caused more degradation than missing facts.

Original post →

More from coding & agent

coding & agent channel →