Most agents start every session from zero. Ori is an open-source, markdown-native memory layer that fixes this. local-first, no cloud, no vendor lock-in.

- Three-signal retrieval: semantic embeddings + BM25 + PageRank fused via RRF

- Graph-aware forgetting: ACT-R vitality model + spreading activation +

Tarjan's articulation point protection (notes that matter stay alive, their neighbors stay warm, the rest fades)

  - Agent identity: name, goals, methodology persist across sessions and clients
 - ~850 tokens per query regardless of vault size
  

- 14 MCP tools, Apache-2.0, 396 tests

  npm install -g ori-memory

  https://github.com/aayoawoyemi/Ori-Mnemos