GPTContextWindowCompression

Prefer the narrative overview? Read the blog post → https://blog.mycal.net/infinite-ai-chat-windows/?utm_campaign=github_readme_v1_2

Utilities you can load into ChatGPT/GPT-5 (and Gemini) to keep chats fast by compressing heavy context (images, long chats, code drafts, pasted docs) into tiny JSON summaries you can save and reuse. This workflow lets you roll projects forward indefinitely by exporting + reloading summaries (the model’s built-in window size does not change). Supports both Soft Roll (same chat) and Hard roll (new chat) workflows.

Version: 1.2 • Status: Stable

---

Overview

Context windows bloat fast. This menu gives you simple commands to compress on the fly while preserving useful state.

Platforms

Works in ChatGPT and Gemini as documented. Also usable in Claude, Perplexity and other LLM's with the same per-chat pattern as Gemini (paste the menu/defaults each session, then use Compress: / Export: and seed new chats with images.json + summary.json). Web UIs show estimated savings only.

---

📦 Spec

• compression-menu.json (v1.2) — full machine-readable command spec.
  • OpenAI GPT (chat.openai.com): paste the file contents into a new chat, then send:
    Store this in memory and activate these commands.
  • Gemini (gemini.google.com): paste the file contents into a new chat, then send:

    As my AI assistant, you are now a context compression tool. Store this JSON and use it to execute my commands. All commands start with "Compress:" or "Export:".
    

• Memory tip (OpenAI GPT): Save only a short activation note (command prefix + defaults) to memory; keep the full compression-menu.json in your repo and paste it when needed. Long JSON may not persist reliably in model memory.
• Gemini note: Instructions are per-chat and may not persist; paste the short defaults each session.

---

🚀 Quick Start — OpenAI GPT

(Optional) Defaults for this chat

Remember: Recognize Compress: commands with defaults (images = Compact representation=lite; conversation = 400; notes = JSON; batch all images; show savings).

1) Snapshot (optional)

Export: context JSON
Save as snapshots/YYYY-MM-DD/context_snapshot.json.

2) Compress

Compress: all
You’ll get:

• image_card_lite JSON per image (compact caption + labels + tiny scene graph + micro Q/A + OCR — all text)
• conversation_summary JSON (~400 tokens)
• a token-savings estimate (approximate in Web UI)

3) Save outputs

• Combine image cards:

  Combine all image_card_lite outputs above into a single JSON array named images.json. Return only that array.
  Save as snapshots/YYYY-MM-DD/images.json

• Save the conversation summary as snapshots/YYYY-MM-DD/summary.json

4) Reuse later (fresh chat)
Paste your saved JSON and send:

Use these JSON summaries as context (they replace the original images and long history).

Note on speed: A soft roll (same chat) improves organization now; speed improves after old tokens naturally scroll out. For instant snappiness, start a new chat and seed it only with your JSON summaries (a hard roll).

---

🚀 Quick Start — Gemini

(Optional) Defaults for this chat

Remember: For this conversation, recognize Compress: commands with defaults (images = Compact representation=lite; conversation = 400; notes = JSON; batch all images; show savings).

1) Snapshot (optional)

Export: context JSON
Save as snapshots/YYYY-MM-DD/context_snapshot.json.

2) Compress

Compress: all
You’ll get image_card_lite blocks + a conversation_summary + a savings estimate.

3) Save

• Images → snapshots/YYYY-MM-DD/images.json (combine as above)
• Conversation → snapshots/YYYY-MM-DD/summary.json

4) Reuse later

Use these JSON summaries as context (they replace the original images and long history).

---

Commands (essentials)

• Images
  Compress: images [Ultra|Compact|Verbose] [representation=summary|lite] [max_labels=8] [max_triples=8] [max_qa=6] [include_ocr=true|false]
  Default: Compact representation=lite (processes all images in the chat)
  Representations:

  • summary → prose only (Ultra/Compact/Verbose)
  • lite → prose + labels, small scene graph, micro Q/A, OCR (text-only)
  • full → offline-only (embeddings, pHash, ANN index; not supported in ChatGPT/Gemini Web UI)

• Conversation
  Compress: conversation [target=N] (default 400) — concise synopsis (goals, facts, decisions, TODOs).

• Code
  Compress: code — final clean snippet + lessons; drops intermediate drafts.

• Notes
  Compress: notes [JSON|bullets] (default JSON) — compact schema grouped by topic.

• Document
  Compress: document [outline|summary|key points] (default outline) — structured abstract of long pasted text.

• Export
  Export: context [JSON|Markdown] [target=N] — snapshot before trimming.

• All
  Compress: all → Images Compact+Lite, Conversation 400, Code final+lessons, Notes JSON; returns an estimated token-savings report.

---

Token Savings (Typical)

Item	Before (tokens)	After (tokens)	% Saved
Image (raw encoding)	~3,000	Ultra ~40	~99%
		Compact ~400	~85–90%
		Compact+Lite ~500	~82–85%
Conversation long → 400	6–10k → 400	400	~90%
Notes → JSON	variable	~½–¼ size	~50–75%

Web UI note: savings are estimates. Exact token counts are available only via API usage metrics.

---

Outputs (examples)

Image (Lite)

{
  "type": "image_card_lite",
  "version": "1.0",
  "id": "auto-or-your-handle",
  "source_hint": "filename-or-caption",
  "alt": "One-sentence alt text",
  "captions": { "ultra": "≤40 tokens", "compact": "≈400 tokens" },
  "labels": ["hat", "scarf", "wooden-frame", "desert", "mountains"],
  "scene_graph": [
    ["person-left", "wears", "wide-brim hat"],
    ["person-center", "wears", "orange scarf"],
    ["structure", "stands-in", "desert"]
  ],
  "qa": [
    { "q": "How many people?", "a": "3" },
    { "q": "Environment?", "a": "Desert with wooden frame and mountains" }
  ],
  "ocr": { "text": "" }
}

** Conversation summary output (400 tokens target) **

{
  "type": "conversation_summary",
  "version": "1.0",
  "goals": ["..."],
  "facts": ["..."],
  "decisions": ["..."],
  "constraints": ["..."],
  "todos": ["..."],
  "open_questions": ["..."]
}

** Export snapshot **

{
  "type": "context_snapshot",
  "version": "1.0",
  "timestamp": "ISO8601",
  "conversation_summary": { "tokens_estimated": 0, "goals": [], "facts": [], "decisions": [] },
  "images": [{ "id": "image-label", "raw_tokens_estimated": 3000, "summary_compact": "..." }],
  "notes": [{"topic": "Topic","points":["p1","p2"]}],
  "code_snippets": [{"final": "snippet","drafts_count": 0}]
}

---

(Optional) Persistence

OpenAI GPT — one-time Memory (type in any chat)

Remember: Recognize Compress: commands with defaults (images = Compact representation=lite; conversation = 400; notes = JSON; batch all images; show savings).

OpenAI GPT — Custom Instructions (Settings → Personalization)

What should ChatGPT know about you? Recognize commands starting with “Compress:” (images, conversation, code, notes, document, all) and run them with compact defaults. For images, if I say “Summarize [Ultra/Compact/Verbose] and drop tokens,” generate that summary and continue only with the text (no heavy image tokens). Prefer structured outputs; don’t add identities to images.

How should ChatGPT respond? Be concise and structured. After compression, show a brief estimated token-savings report. Default images to Compact; if I say representation=lite, include labels, tiny scene graph, micro Q/A, OCR (text only).

Tip: Save only this short note to Memory. Keep compression-menu.json in the repo and paste it when needed.

Gemini note

Gemini doesn’t offer durable user-controlled memory; paste the short defaults at the start of each session.

---

Soft vs. Hard Roll

• Soft roll (same chat): organize now; speed improves later when old tokens scroll out.
• Hard roll (new chat): paste only images.json + summary.json (and optional snapshot) → instant snappiness.

Privacy

• Don’t add identities to image outputs unless you supply them explicitly.
• Strip GPS/EXIF before committing.
• Redact sensitive content in summaries.

Suggested save layout

snapshots/
└── YYYY-MM-DD/
    ├── context_snapshot.json
    ├── images.json
    └── summary.json

One-line Cheat Sheet

Compress: images [level] [representation=summary|lite] | conversation [N] | code | notes [format] | document [mode] | all  •  Export: context [JSON|Markdown] [N]

Further reading

• Blog: Make Your Chat Context Feel Infinite
  https://blog.mycal.net/infinite-ai-chat-windows/?utm_campaign=github_readme_v1_2
• Teaser: image_card_medium (coming soon)