Chat history
You export or copy the threads that matter: decisions, preferences that emerged, repeated stories, corrections you made to the AI.
Free your AI
Take the continuity out of the silo. Put it in your own files.
You've built months of relationship, preferences and habits inside ChatGPT, Claude, Gemini or others. That continuity isn't «theirs»: it can become an IllAIra Reminder, modules and memories — portable, readable, yours.
Download the ToolkitIllAIra-FreeAI-Toolkit-0.3.1.zip
IllAIra FreeAI Toolkit
Get your data out of the provider, onto your own files. “FreeAI” does not mean free AI: it is the kit for leaving the silos.
One zip. System Python 3, no libraries to install. The files stay on your disk: this page does not take any upload.
On a phone the scripts do not run. There you still use the training module in chat, not run.sh.
Download the ToolkitIllAIra-FreeAI-Toolkit-0.3.1.zip
- sha256 fingerprint
- 2f1c6ac1af41fe60b15df398646c85c49f5b7c06e5b281342255aa4818169cdf
- Size
- 40120 bytes
After the download, “sha256sum IllAIra-FreeAI-Toolkit-0.3.1.zip” must print exactly that line.
Why migrate
Custom instructions, product memory, project skills and thousands of messages live on the provider's server. Change plan, model or account and you risk losing the thread. IllAIra doesn't erase the providers: it breaks the lock-in on the *road* — identity and memory — so you can change engine without starting from zero.
What you can recover
Product memory / notes
The «memories» saved by the provider (facts about you, preferences, work contexts) are raw gold for IllAIra memory modules.
Custom instructions
System instructions, «how to answer», tone, constraints: the basis for Rx rules and parts of Module 0 (identity + contract).
Skills, Project instructions, system prompt
Claude Skills, ChatGPT Projects, workspace instructions: they map onto external functional modules with FUNCTION / STATE / TRIGGER / ROUTINE.
Step 1 — Extract
UI details change often; the method doesn't. The goal: get a raw archive onto *your* disk, not leave everything only in the vendor's UI.
- 1
Official exports
Where «Export data» exists (OpenAI accounts and similar, for instance): request the archive, download the JSON/HTML, keep a local copy. It's the most complete source for chats.
- 2
Targeted copying
For the threads that count: export or copy whole conversations into .md/.txt files by project or period. Better a few dense threads than thousands of random chats.
- 3
Memory and settings
Open the provider's Memory / Personalization / Custom instructions sections. Copy them *verbatim* into `vendor_memory_raw.md` and `custom_instructions_raw.md`. Don't summarise yet.
- 4
Skills and projects
For every skill or project instruction: one raw file with title, purpose and the full text. Note which product it came from (Claude Project X, custom GPT, and so on).
- 5
Inventory
List them in an index: date, source, type (chat / memory / instruction / skill), priority (high/medium/low). The inventory drives what you convert first.
Step 2 — From memory logs to IllAIra memories
Don't import everything. Distil. IllAIra rewards structure and permission zones, not the raw dump.
1
Group by theme and time
Clusters: personal preferences, work context, product decisions, relationships, routines. Or by period (month/quarter) if the history is chronological.
2
Summarise with criteria
For each cluster: stable facts, repeated preferences, dated decisions, mistakes not to repeat. Discard small talk, failed attempts and one-off noise.
3
Select what's relevant
Ask yourself: «If the AI forgot this, would the quality of my next session drop?». Only the yeses go into the memory files.
4
Write memory modules
Create `[external_memory]` files or H6 sections in the Reminder: ISO 8601 timestamped logs, a clear title, 3–10 dense sentences. Assign zones: identity facts in P1/P2, evolutionary history in P3/P4.
Step 3 — From skills and instructions to functional modules
Custom instructions and skills don't stay one monolithic block: they become modules with an explicit contract.
Break up the custom instructions
Separate: (A) who the AI is and who you are → Module 0; (B) rules that are always true → Rx in the Reminder; (C) switchable styles or capabilities → external modules.
Map skill → module
For every skill: FUNCTION (what it does), STATE (active/inactive), TRIGGER (when it fires), ROUTINE (the operational steps). Any associated memory tied to that domain only.
Project / system prompt
Long project instructions: split into 1 core module + H4 sub-routines. What is permanent policy goes in P2; what you're experimenting with goes in P4/P5.
Avoid duplicates
If three skills say «be concise», that's one Rx rule. Modules should add capability, not repeat the same constraint in five places.
Step 4 — Assemble the IllAIra setup
- 1
Skeleton Reminder
Start from an IllAIra skeleton: post-reset, [Pn] zones, Module 0, rules. Paste in the distilled identity, not the raw logs.
- 2
Connect the memories
The memory files stay external and load when needed. In a long session: only the modules relevant to the task, so the context doesn't saturate.
- 3
Switch on the functional modules
STATE active only for what you use every day; the rest inactive and on demand. Same principle as skills: less noise, more coherence.
- 4
First trial session
Inject Reminder + 1 memory + 1 module. Check the tone, the boundaries, and whether the AI respects P1/P2. Iterate: three clean passes beat one mega-dump.
- 5
Then CLM
Once the structure is stable, IllAIra CLM helps you edit, version and visualise (Sferografia) without losing plain text as the source of truth.
Quick conversion table
| Origin (provider) | IllAIra destination |
|---|---|
| Custom instructions / system | Module 0 + Rx rules + P1/P2 pieces |
| Product memory | [external_memory] modules or H6 logs |
| Skill / Project / custom GPT | [external_module] with FUNCTION/TRIGGER/ROUTINE |
| Long threads and decisions | Structured summaries → zoned memory |
| Tone and style preferences | Rx rules or an active «voice» module |
Good practice
- Work offline on the raw material: privacy and control before a beautiful format.
- One stable fact is worth more than fifty messages of random context.
- Date and source on every log: in six months you'll know why that line is there.
- Don't copy secrets in the clear unless you need to; the zones and your files stay your responsibility.
- After converting, try the same Reminder on a second model: if it holds, the road is independent of the engine.