LISA - AI Chat Exporter for Claude, ChatGPT & More

Your context, your choice of AI, no starting over. 🎯 WHAT IS LISA? A privacy-first browser extension that captures, compresses, and preserves your AI conversations — so you never lose context, never start over, and never get locked into one platform. LISA stands for Linguistic Intelligence Semantic Anchoring — structured data that pre-resolves ambiguity before AI inference begins. 🔒 THE PROBLEM Your AI conversations are trapped. ChatGPT won't let you continue in Claude. Gemini doesn't know what you discussed in Grok. Every platform locks your data in its own silo. No export. No portability. No ownership. You created that knowledge. You should own it. ✨ THE SOLUTION LISA extracts conversations into structured files that work across any AI platform. Inject from your library or upload to Claude, ChatGPT, Gemini, Grok, or any other assistant supported and instantly restore context — including code blocks, decisions, and reasoning. Never Re-Explain Your Project to an AI Again Install LISA, open any AI chat, click capture, and hand the file to a different AI. The first time it picks up exactly where you left off, you'll get it instantly. All new users get 100 free compressions. 🔹🔹🔹 ⚡ NEW — v0.52.9: - Full support for Claude in Chrome conversations. - Active Context Management (ACM): keep long AI conversations sharp, and take them anywhere. More details below. ⚡ 🔹🔹🔹 ✨ LISA's Core Principle: The compression pipeline is essentially a translator. It translates a human conversation into a machine-readable signal. And like any translation, the question is: what carries meaning and what's just grammar holding the sentence together? Noise falls into three categories: Structural glue — "the", "and", "with", "also", "just". These words connect ideas in human language but carry zero semantic signal. An AI reconstructing context from "tokens lean lisa conversation fix session" understands exactly the same thing as "the tokens in the lean lisa conversation to fix the session". The glue is for humans, not machines. Quantifiers without anchors — "all", "two", "some", "real", "many". These modify something but alone they point to nothing. "Two" only has meaning next to what it counts — "two edits" has signal, "two" alone is noise. The concept graph captures "two" as a standalone node because it appears frequently, but it tells the receiver nothing. Narrative scaffolding — "verified", "noticed", "showed", "because". These describe the act of discovering something, not the thing itself. "Never verified because you didn't reload" is a story about what happened. "Always use getattr() on Stripe objects" is a rule. The receiver needs rules, not stories. Signal is the opposite — words that are definite, specific, and independently meaningful: Domain nouns: "tokens", "compression", "anchor", "harness" — each one names a thing Action verbs in context: "fix", "strip", "merge" — each one names what happened Named entities: "LISA", "SemanticAnalyzer", "Stripe" — each one names who/what When you stack signal words together — "tokens lean lisa compression fix session" — they form a semantic skeleton that an AI can expand back into full understanding. The stopword stripping isn't destroying information, it's removing redundancy that the receiver will regenerate automatically from context. This is literally what you do as a translator. You don't translate word-for-word from Arabic to English — you translate meaning. Some Arabic particles have no English equivalent and get dropped. Some English articles have no Arabic equivalent and get added. The meaning crosses, the grammar adapts. That's what LISA does between human language and machine language. The deeper insight: the frequency filter catches what the static list can't. Any word appearing in >25% of sentences is probably glue for this specific conversation, even if it's a real noun in other contexts. "File" is meaningful in a cooking conversation, noise in a coding session where every other sentence mentions files. AI doesn't read words. It reads tokens — subword fragments produced by a tokenizer. "Understanding" is one token. "ACME-Tech Co." is three. Every token costs compute, and raw conversation is full of waste: "sure", "got it", "thanks" burn tokens that carry zero meaning. Pronouns like "he", "it", "the tool" force the AI to spend inference work guessing what they refer to. The same entity name repeated forty times costs forty times the tokens. And none of this ambiguity was yours — it's an artifact of how conversation flows, not what you meant. LISA's compression pipeline attacks each layer of that waste. Salience routing scores every turn by how central it is to the conversation's meaning and drops the noise. Entity aliasing replaces repeated names with short aliases backed by a glossary — "ACME-Tech Co." appears once in full, then as "ATC" for the rest, with zero ambiguity. Coreference resolution rewrites "he said it was done" as "Amar said the parser was done" before the AI ever sees it. And real token counts (via tiktoken, the same tokenizer the models use) replace character-based guesses so you know exactly what you're saving. The result: the AI gets a pre-disambiguated, leaner input where every token carries meaning. ⚡ v0.52.9: Full support for Claude in Chrome conversations. Conversations you have with the Claude in Chrome extension now open on claude.ai with LISA working as usual. LISA reads them through the same channel claude.ai’s own page uses, so it captures every message, from the first to the last, however long the session ran. Export, Compress, the Markdown view, the context monitor and Handoff all see the complete conversation. Claude’s browser actions (navigating, reading pages, running scripts) are recorded as tool steps, so your summaries stay focused on what was said. Normal claude.ai chats work exactly…