MemPalace: 170 Tokens to Recall Everything — A Long-Term Memory System for AI Agents

Originally published at recca0120.github.io Six months of daily AI conversations. 19.5 million tokens of history. Start a new session and it remembers nothing. You can dump important things into CLAUDE.md, but that file quickly balloons to thousands of lines, eating up your context window on every startup. MemPalace takes a different approach: instead of cramming all memories into the prompt, build a structured memory vault that AI queries on demand. Startup loads just 170 tokens, search accuracy hits 96.6%, completely offline, zero API calls. The Memory Palace Architecture MemPalace uses the ancient Greek memory technique as its organizational metaphor: Wing : Projects, people, or topics. One wing per major category Room : Sub-topics within a wing — auth, billing, deploy Hall : Memory type corridors shared across all wings hall_facts — locked-in decisions hall_events — sessions and milestones hall_discoveries — breakthroughs hall_preferences — habits and opinions hall_advice — recomme