billion-context Rewrites Long Coding Sessions.
billion-context is an open-source proxy that incrementally compresses coding-agent history into reversible, model-generated summaries while keeping recent work raw. Its paper reports multi-month sessions, billions of cumulative tokens, and zero context-window violations in production logs. [GitHub repository](https://github.com/ranxianglei/billion-context)
The compelling idea is not a larger context window, but a smarter memory policy that preserves lineage instead of burying history in one irreversible summary.
- –Works across Anthropic and OpenAI-compatible API formats, making it applicable to multiple coding-agent hosts.
- –Decompression, search, and status tools give agents more control than threshold-triggered automatic compaction.
- –The project reports strong longitudinal results, but its headline 5.3× savings figure is identified as analytical or counterfactual rather than fully validated.
- –Prefix-cache preservation could make compression materially cheaper than repeatedly rewriting entire histories.
- –Developers should test recovery quality on their own repositories before trusting it with high-stakes, long-running sessions.
DISCOVERED
1h ago
2026-11-06
PUBLISHED
1h ago
2026-11-06
RELEVANCE
AUTHOR
AICodeKing