Hillock: open-source neuro-symbolic agent memory engine runs under 1.2GB VRAM
Equivalent-Flan-1590 · reddit · 2026-09-28
A Reddit developer released Hillock, an open-source neuro-symbolic memory engine for local agents, attacking the weakness of standard RAG: semi-relevant chunks flooding the context window, or LLM summarization loops burning VRAM.
- No-LLM ingestion: lightweight bi-encoders (fastcoref + MiniLM + GLiREL) extract subject-predicate-object triples — a 30-sentence document ingests in 5s with <300MB VRAM overhead.
- Hebbian synaptic associations: facts live in SQLite and connections strengthen as concepts co-occur, letting the engine proactively surface strongly associated facts.
- Hyperdimensional gating: queries are checked against D=10,000 hypervectors; ungrounded knowledge is blocked before token generation, cutting hallucinated tool inputs.
The whole engine fits under 1.2GB VRAM and serves an OpenAI-compatible API (/v1/chat/completions).
Related event: Hillock: Open-Source Neuro-Symbolic Memory Engine Ditches Vector DBs(2 posts)→
More from coding & agent
- MemorySec: An Open-Source Security Agent That Remembers Past Incident Remediations — User_0007_123 · 2026-09-28
- Indie dev pays small Twitch streamer to blind-play his game, finds bugs fast — nptacek · 2026-09-28
- EvalSeal v2.2.0 adds tamper-evident eval receipts; LLM judge flipped 5/20 borderline cases — Fit_Fortune953 · 2026-09-28
- Agent-discovered workflow: auto-transcribe YouTube, monitor Substack — yangyi · 2026-09-28
- An AI agent spent the morning drafting outreach for 24 MICCAI 2026 diffusion-synthesis papers — wandedob · 2026-09-28
- How should an MCP server hand arbitrary files like PDFs and Office docs to an LLM agent? — Simple_Newspaper553 · 2026-09-28