DUMEdb: open-source engine replaces 100k-token context replay with 1959-style unit-record memory
MadBradSmith · reddit · 2026-10-20
A developer open-sourced DUMEdb (Deterministic Unit-Record Memory Engine), targeting the "session amnesia" pain point of long-running agents: every restart replays 80k–120k tokens of transcripts, burning money, slowing local inference, and causing hallucinations past the 75% context ceiling.
The design borrows immutable unit-record accounting from the 1959 IBM 1401: instead of re-reading conversations, memory is stored as atomic fixed-column registers. Key mechanics:
- Zero-turn wake handshake (cabinetWake): the agent passes a codeword and gets a status receipt plus active registers in under 220 tokens, preserving 99.8% clean context
- Deterministic LMDB engine: memory-mapped storage with sub-microsecond zero-copy register reads
- Disk odometer: audits transcript.jsonl directly for ground-truth token consumption and triggers clean session handoffs before auto-compaction amnesia
- Native MCP bridge: works with Claude Desktop, Cursor, Google Antigravity, and Python agent loops
Available via pip install dumedb, with a GitHub repo, live demo, odometer calculator, and an 88-second promo video.
More from coding & agent
- Agent framework roundup: CrewAI runs in production at Fortune 500 firms — mdancho84 · 2026-10-20
- LlamaIndex, the 52.6k-star RAG framework, tops AI engineer job descriptions — mdancho84 · 2026-10-20
- LangGraph, the 43.6k-star stateful agent orchestration framework used by Klarna and Replit — mdancho84 · 2026-10-20
- 10 GitHub repos to become an AI engineer in 90 days, from AI Agents for Beginners to Qdrant — mdancho84 · 2026-10-20
- Harness Lens compares 19 coding agent harnesses: system prompts and compaction exposed — LysandreJik · 2026-10-20
- "Vibe coded pieces collapse": theo argues engineers are still very much necessary — karmay007 · 2026-10-20