MemPalace stores every AI conversation verbatim, then makes them instantly findable. 96.6% on LongMemEval with zero API calls — beating every paid competitor. Built by Milla Jovovich & Ben Sigman using Claude Code.
pip install mempalace
Decisions happen in conversations now — not in docs, not in Jira. In conversations with Claude, ChatGPT, Copilot. Six months of daily AI use equals 19.5 million tokens of reasoning, decisions, and breakthroughs. All trapped in chat windows that evaporate when the session ends.
| Approach | Tokens Loaded | Annual Cost |
|---|---|---|
| Paste everything into context | 19.5M — doesn't fit | Impossible |
| LLM summarization tools | ~650K tokens | ~$507/yr |
| ⭐ MemPalace wake-up (L0+L1) | ~170 tokens | ~$0.70/yr |
| ⭐ MemPalace + 5 searches | ~13,500 tokens | ~$10/yr |
Unlike Mem0 and Zep — which use an LLM to decide what's "worth remembering" and discard the rest — MemPalace stores your actual exchanges in ChromaDB without summarization or extraction. The 96.6% LongMemEval result comes from this raw verbatim mode.
Inspired by the ancient method of loci used by Greek orators. Your conversations are organized into Wings (people/projects), Rooms (topics), Halls (memory types), Closets (summaries), and Drawers (verbatim originals).
Every conversation stored in ChromaDB exactly as written — no LLM deciding what to throw away. Semantic vector search retrieves the right memory in milliseconds. This is what drives the benchmark-topping 96.6% score.
Halls connect rooms within the same wing by memory type: facts, events, discoveries, preferences, advice. Tunnels cross-reference the same topic across different wings automatically.
L0 (identity, ~50 tokens), L1 (critical facts, ~120 tokens), L2 (room recall, on demand), L3 (deep semantic search, on demand). Your AI wakes up with 170 tokens and knows your entire world.
Temporal entity-relationship triples stored in SQLite — free alternative to Zep's Graphiti (Neo4j). Track who decided what, when, and why. Invalidate outdated facts. Query any point in the past.
Create agents that focus on specific domains — code reviewer, architect, ops. Each gets its own wing and AAAK-compressed diary. Scale to 50+ agents with zero CLAUDE.md bloat. Replaces Letta ($20–200/mo) at zero cost.
┌─────────────────────────────────────────────────────────────┐ │ WING: wing_kai (Person) │ │ │ │ ┌──────────────┐ ─hall─ ┌──────────────┐ │ │ │ hall_facts │ │ hall_events │ │ │ └──────┬───────┘ └──────────────┘ │ │ │ room: auth-migration │ │ ▼ │ │ ┌──────────┐ ┌─────────────────────────────┐ │ │ │ Closet │─────▶│ Drawer (verbatim original) │ │ │ └──────────┘ └─────────────────────────────┘ │ └───────────┼─────────────────────────────────────────────────┘ │ tunnel (same room topic, different wing) │ ┌───────────┼─────────────────────────────────────────────────┐ │ WING: wing_driftwood (Project) │ │ │ room: auth-migration │ │ ┌──────┴───────┐ ─hall─ ┌──────────────┐ │ │ │ hall_facts │ │hall_discoveries│ │ │ └──────┬───────┘ └──────────────┘ │ │ ▼ │ │ ┌──────────┐ ┌─────────────────────────────┐ │ │ │ Closet │─────▶│ Drawer (verbatim original) │ │ │ └──────────┘ └─────────────────────────────┘ │ └─────────────────────────────────────────────────────────────┘ Halls: hall_facts · hall_events · hall_discoveries · hall_preferences · hall_advice Search impact: All closets: 60.9% R@10 → Wing+Room filtered: 94.8% R@10 (+34%)
MemPalace raw mode (zero API, zero cost) outperforms every paid competitor on the industry-standard LongMemEval benchmark.
| System | LongMemEval R@5 | API Required | Monthly Cost | Local |
|---|---|---|---|---|
| ⭐ MemPalace hybrid (Haiku rerank) | 100% | Optional | Free | ✅ |
| ⭐ MemPalace raw (zero API) | 96.6% | None | Free | ✅ |
| Supermemory ASMR | ~99% | Yes | Paid | ❌ |
| Mastra | 94.87% | Yes (GPT) | API costs | ❌ |
| Mem0 | ~85% | Yes | $19–249/mo | ❌ |
| Zep | ~85% | Yes | $25/mo+ | ❌ |
# Install pip install mempalace # Initialise your palace (guided, detects people & projects) mempalace init ~/projects/myapp # Mine your conversations (Claude, ChatGPT, Slack exports) mempalace mine ~/chats/ --mode convos # Search six months of context in milliseconds mempalace search "why did we switch to GraphQL" → "Chose GraphQL over REST — concurrent writes, dataset exceeds 10GB. 2025-11-03"
"We tested MemPalace at The Zero-Human Company and deployed it to 79 employees. This is a masterpiece."
Brian Roemmele — AI Industry KOL"Milla Jovovich launching an AI memory system was NOT on my 2026 bingo card."
Wayne Sutton — 1.2M+ impressions"Milla Jovovich has a GitHub. She co-developed the highest-scoring AI memory system. What a boss."
@am_will — Twitter community"Yes, that Milla Jovovich. Missed opportunity to call it Resident Eval." — top HackerNews comment, denysvitali
HackerNews — Top comment on launch day