Actress & Dev Beat VC-Backed AI Memory Tools—For Free

Your AI coding assistant just helped you debug a gnarly race condition. Three days later, you're back in the same codebase—and the conversation history is gone. Every architectural decision, every why-not-this-approach discussion, vanished. The model can process millions of tokens, but it can't remember last Tuesday.

Two developers just shipped a fix for that. And they're giving it away.

The Problem: Millions of Tokens, Zero Memory

AI conversation history disappears after each session, erasing months of debugging sessions and architectural debates. Claude Code, ChatGPT, Cursor—all can process context windows that could swallow entire codebases, but none remember what you talked about yesterday.

Commercial solutions like Mem0 and Zep tackle this by extracting facts and summarizing conversations. Store the insights, discard the raw material. It's lossy compression for dialogue—fast, efficient, and works well for many use cases.

MemPalace does the opposite. It stores verbatim content and uses semantic search to retrieve what you actually said, not a distilled version. No summarization layer. No extraction pipeline. Just your conversations, searchable.

The Underdog Team: Claude Code and Unconventional Collaboration

The project came from Milla Jovovich—yes, the actress—and developer Ben Sigman, built using Claude Code as an AI-assisted coding tool. It's a strange origin story that doubles as a proof point: modern AI development tools can level the playing field against well-funded engineering teams.

This isn't about celebrity novelty. It's about two people shipping something competitive with teams that have raised millions.

Verbatim vs. Extraction: A Philosophical Fork

Mem0 and Zep extract and summarize because it handles scale well, stores metadata efficiently, and works when you need the gist of past conversations without the verbatim text.

MemPalace makes a different trade-off. By keeping full conversations and using ChromaDB for vector storage, it bets that developers want the exact words—the specific phrasing of a decision, the precise error message from two weeks ago, the context around why a particular approach was abandoned.

Both approaches solve real problems. One optimizes for storage and speed. The other optimizes for fidelity and retrieval precision.

The Benchmark Controversy: 96.6% vs. 100%

The project initially claimed a 100% score on LongMemEval. That number came with an asterisk: it was achieved by hand-tuning fixes for specific failing questions and re-testing on the same questions—a benchmark integrity violation worth acknowledging.

Strip away the tuning, and the score lands at 96.6%. Still ahead of Mem0's ~85% and Zep's ~82%, but without the perfect-score headline. The controversy doesn't erase the achievement, but it matters for anyone comparing numbers.

The experimental AAAK compression feature—designed to abbreviate stored text—currently drops performance from 96.6% to 84.2%, so it's worth leaving disabled for now.

MCP Integration: Works Today, No Migration

MemPalace ships with 19 MCP tools that integrate with Claude Code, ChatGPT, Cursor, and Gemini CLI. The Model Context Protocol support means you can drop it into your existing workflow without rewriting scripts or migrating data.

The memory stack loads 170 tokens at session start—critical facts from Layer 0 and Layer 1—then performs on-demand semantic searches for deeper queries. Layer 2 pulls room-specific memories; Layer 3 runs full-depth searches when you need something from six months ago.

Storage runs on ChromaDB for vectors and SQLite for metadata. Self-hosted, local, no cloud dependency unless you want one.

What You Can Mine: Slack, Transcripts, Local Files

The project supports mining Slack exports, Claude and ChatGPT conversation transcripts, and local project files. Entity and relationship detection runs automatically, building a knowledge graph of people, projects, and decisions as you import data.

You're not starting from scratch. Export your ChatGPT history, point the importer at it, and your past conversations become searchable context.

Open Source vs. Subscription: The Real Competition

Mem0 charges $19–$249/month. Zep starts at $25/month. Both offer cloud-hosted infrastructure, managed scaling, and professional support. They've built serious engineering around real problems.

MemPalace costs $0 and runs on your machine. It's not better or worse—it's a different model. The commercial tools trade cost for convenience and uptime guarantees. MemPalace trades setup effort for full control and zero recurring fees.

Where enterprise tools optimize for cloud infrastructure and managed services, this focuses on local control and verbatim storage. The open-source approach expands choice.

The GitHub repo has installation docs. Try verbatim memory. See if lossless beats extraction for your workflow.