Give your AI Agent
long-term memory

Why NylonME

Memory is not retrieval — it is resonance

Most memory systems treat memories as flat text chunks fetched by top-k similarity. NylonME weaves them into a living network: understood at write time, reflected on when idle, and mutually reinforcing along edges at read time.

Integrations

Five ways into your agent

Embedded on one machine, zero-binary over HTTP, a shared bridge across machines, straight SDK calls, or right inside a phone app — one engine, one write path.

On-Device

The same engine, in your phone

nylon-mobile compiles the exact same weave / resonate / retrieve implementation (the shared nylon-service core) into Android / iOS apps — no daemon, no port, no network calls. Memories never leave the device. UniFFI generates the Kotlin / Swift bindings.

Optional on-device semantics: a quantized bge-small-zh ONNX model with pure-Rust tract inference — a zero-lexical-overlap probe like "do they have a pet at home" still lands on the right memory. WAL-first writes plus automatic checkpoints (every 32 writes / 300s by default) survive process kills; two lifecycle hooks (onStop → checkpoint, onDestroy → shutdown) cover the rest. Details in Docs · Mobile SDK.

Quick Start

One MCP config, and your agent remembers

nylon-engine mcp speaks Model Context Protocol over stdio with the engine embedded in-process — no daemon, no port, data lands in ~/.nylonme/data automatically.

{
  "mcpServers": {
    "nylonme": {
      "command": "/path/to/nylon-engine",
      "args": ["mcp"],
      "env": { "NYLON_OWNER": "my-project" }
    }
  }
}

Want to see it running first? No Docker required — Windows users can grab an archive from Releases and just unzip it; with Docker, one command also brings up the ollama embedding model:

# Option 1 (recommended): no Docker needed — Windows / macOS / Linux
# Download nylonme-windows-x64.zip from GitHub Releases, unzip, then:
nylon-engine.exe serve 127.0.0.1:50051
start http://127.0.0.1:50052
# No Rust toolchain, no key — a single machine runs in open mode out of the box

# Option 2: Docker (also brings up ollama embeddings for semantic recall)
git clone https://github.com/nylon-memory/NylonME.git
cd NylonME
docker compose -f docker-compose.prebuilt.yml up -d
# gRPC on :50051 · web console / REST / /mcp all on :50052
# open http://localhost:50052 to browse your first memories

LoCoMo Benchmark

86.0% — earned one explainable step at a time

Full 10-session corpus, 1536 answerable QAs, lexical + vector fused retrieval. Every layer of the experiment delivered a lift we can explain.

Per-category recall: multi-hop 82.3% · temporal 89.1% · single-hop 89.1% · commonsense 58.7%. End-to-end QA (LLM answers from retrieved evidence, judge-graded): paper metric 82.9% — same weave store and answer model throughout: 76.4% → anti-refusal prompt 80.1% → widened answer context 82.9%.

Second benchmark LongMemEval-S (100-instance slice): evidence recall any-hit 99%, end-to-end 83.0%. Widening the answer context from Top-10 to Top-15 gained +5pp, and the same mechanism reproduced +2.8pp on LoCoMo. Full methodology lives in the two benchmark reports under docs/ in the repo.

Roadmap

From runnable to widely adoptable