Memory is what makes an agent an agent
Most assistants forget the moment the transcript scrolls away. NMAFC is an open-source memory layer that keeps a dated, decay-aware, source-attributed record of everything it learns — and hands back a small, precise context hundreds of sessions later.
NMAFC — memory that outlives the context window
Most assistants forget. Not because the model is weak, but because the transcript ran out. NMAFC is an open-source memory layer that compacts every turn into dated atomic facts, ages unused ones on a decay curve set for the long haul — at the shipped λ = 0.005 a fact takes roughly 460 turns of neglect to demote, so nothing decays mid-conversation — and gives back a small, precise, source-attributed context, hundreds of sessions later.
It is a Python package with a server attached. Import it, or point an agent at it. No OpenLoft account, no vector database to provision, and no conversation content leaving your network when you run a local model and a local embedder together — the extractor still ships turns to whichever LLM you configure it with. Local embeddings on their own do not close that gap. Read the docs for install steps, configuration and the honest limitations.
Four ways to connect
The OpenAI-compatible endpoint mirrors the chat-completions shape but drops tools and tool_choice, so it is not safe for tool-calling clients — most agent harnesses are. Use MCP or REST for those; the routes are in the docs.
LoCoMo accuracy · 1,985 paired questions
vs 46.1% vanilla RAG · +25.5 · McNemar p ≈ 1e-14
vs 44.8% · +12.5 · p = 0.0075 · under our 800-question floor
vs 45.6% · +5.3 · p = 0.13 — a null result, not a win
vs 50.4% · -14.6 — we lose this one, and publish it
Temporal is the only category whose lead clears significance, and it is listed first for that reason rather than because it is the largest. Scores come from our own LLM judge — no published reliability data and no human spot-check, so do not read “independent” into this panel. The full table across all five categories, including the two where vanilla RAG wins, is in the docs, and every figure regenerates from the committed data with python summarise.py.
Benchmark: LoCoMo (Maharana et al., ACL 2024), used under its CC BY-NC 4.0 licence — non-commercial. These are research results on an academic benchmark, not a shipping claim about a product evaluation.