Heimdall: open-source CPU-only memory system gives coding agents cross-repo memory
Slight-Parfait3679 · reddit · 2026-09-15
Heimdall (MIT, open source) gives AI coding agents persistent memory across all your repositories and projects, so "did I already solve this elsewhere?" becomes one verified search instead of 20 minutes of grep.
Key points:
- Cross-project semantic graph: unlike per-project memory tools, it indexes everything you touch into one semantic graph; integrates with Graphify, Graft, and Hermes
- Zero token spend: maintenance runs as a local daemon (file watching + tree-sitter AST parsing + sqlite); indexing is CPU-only, never an LLM call; retrieval is hybrid lexical+semantic+graph search over local embeddings, with optional GPU for 3.4x speedup
- Self-healing: a single deterministic reconciler prevents agents acting on stale graphs, no full rebuilds needed
Facts and key decisions are stored too. Best LongMemEval S score so far is 0.94, though benchmarks aren't the focus; the project is experimental and seeking contributors.
More from coding & agent
- Tencent open-sources Cube Sandbox: sub-60ms cold-start runtime for AI agents — udmrzn · 2026-09-15
- Team kills AI security startup, open-sources local red-team engine OpenHunterAI — udmrzn · 2026-09-15
- Cline launches Desktop app purpose-built for open-weights models — cpaik · 2026-09-15
- Dan Wahlin: the real superpower is recognizing when AI is wrong, not prompting it — DanWahlin · 2026-09-15
- He spends $14k/month on AI subs and sells agent skills for $20/month — doodlestein · 2026-09-15
- C3 AI's DIA paper accepted at EMNLP 2026, tops all seven SQL benchmarks autonomously — C3_AI · 2026-09-15