Hi HN, FERNme is a memory layer for AI agents. Most systems (Mem0, etc.) call an LLM on every turn to extract/summarize memory, or dump everything into a vector DB. I wanted to see how far a cheaper, more brain-like approach goes. Each user is a node in a fuzzy preference graph; edges (0–9, plus explicit negative edges for dislikes) update with a Hebbian co-occurrence rule — no LLM on the write path — decay ACT-R-style, and compile to a flat, ~40-token "memory card." Retrieval is spreading activation, not vector search. Some early results (all reproducible in the repo): ingesting 86 free-form diary entries about one person ran with 0 write-time LLM calls, kept a flat ~40-token card, and on a LoCoMo-style QA set the context-seeded retrieval answered ~4× more questions than frequency/recency baselines at equal budget — and was the only LLM-free method that handled preference drift. Honest about what this is not: the Hebbian + spreading-activation mechanism isn't novel — HippoRAG, Ori-Mnemos, and HeLa-Mem all use versions of it. My bet is on the combination: near-zero-cost writes, multi-tenant + a privacy-preserving population prior, user-owned cross-surface memory, and evaluating on actions rather than QA. Benchmarks so far are synthetic or single-person; a real LLM-memory (Mem0) head-to-head needs API keys and isn't run yet. It's MCP-compatible, and the repo has a paper draft (PAPER.md) plus a reproducible demo (demo/elena/ — builds memory from 86 free-form diary entries and runs the QA benchmark). Code: github.com/mirkofr/FERNme I'd love feedback on: the memory representation, whether the action-coupled eval is meaningful, failure modes you'd expect, and benchmarks/competitors I should test against — especially a real Mem0 head-to-head. Criticism very welcome.