billion-context plugin compresses billion-token agent sessions into a 100K window
aigclink · x · 2026-10-24
Open-source plugin billion-context (2.5k GitHub stars) argues you don't need a huge context window — let the agent manage context itself. Through continuous compression plus on-demand recovery, it claims a 100K window can support month-long sessions accumulating billions of tokens, using 5x fewer tokens.
Unlike one-shot summarization, history stays recoverable:
- Incremental hierarchical compression: the agent chooses which old messages to fold into summaries
- On-demand recovery: decompress pulls back original content when tracing early decisions or tool calls
- History search: searchcontext finds relevant history in compressed summaries, then restores originals
- Prompt cache optimization: stable cache prefix to cut recomputation cost
Works with Claude Code, Codex, OpenCode, DeepSeek Harness and other coding agents.
Related event: billion-context: Run Billion-Token Sessions in a 100K Window(2 posts)→
More from coding & agent
- Codex users report recurring 'model at capacity' errors breaking overnight agent runs — CtrlAltDwayne · 2026-10-24
- A wearable pendant that orchestrates Claude, GPT, Gemini and more voice-first — EricBuess · 2026-10-24
- Dev warns AI code suggestions are architecturally bad: "like shoving garbage under the bed" — IanArawjo · 2026-10-24
- Nucleus 2026 Low-Code Matrix: Appian and OutSystems Lead as AI Agents Blur Categories — AcanthaceaeLatter684 · 2026-10-24
- MOSS-Image Ecosystem Roundup: Alpha-Channel LoRA, Hi-Res Adapter, A1111-Style Prompts — optimisticalish · 2026-10-24
- Reverse Engineer Anything hits 357k GitHub stars in 6 months by giving AI coding agents a decompiler — chessbuzz · 2026-10-24