For developers
Wire it in this afternoon.
MemMesh sits underneath whatever you already build with, whether that's Claude Code, Cursor, LangGraph, CrewAI or a loop you wrote yourself. Install it over MCP with one command, or call it from an SDK. There is no framework to adopt and no agent abstraction to inherit.
npx @memmesh/cli installThe surface
Four calls, and that's the API you need on day one.
- observe()
- Hand the engine raw text or a typed observation. What it keeps is decided in code, not by asking a model.
- recall()
- Get the context for one subject, sources attached, ready to drop into a prompt.
- search()
- Semantic and keyword retrieval across the graph, limited to what this key is allowed to see.
- predict()
- Name what you want predicted and get a forecast from that subject's history, or an abstention if the evidence is thin.
Quickstart
Pick your language.
Every SDK exposes the same observe, recall and search calls. TypeScript, Python and Go install from their registries today. .NET and Rust are public source while we finish publishing their packages.
npm install @thinkfleet/memory-sdkimport { ThinkFleetMemory } from '@thinkfleet/memory-sdk'
const tf = new ThinkFleetMemory({ apiKey: 'sk-...', projectId: 'proj_...' })
// Remember something
await tf.memory.observe({
subject: { kind: 'contact', externalId: 'sarah' },
content: 'Prefers email over phone.',
})
// Recall it, semantically
const hits = await tf.memory.admin.search({ query: 'how to reach sarah', limit: 5 })
// Synthesize higher-order insights from what it has learned
const { insights } = await tf.memory.admin.reflect({ maxInsights: 3 })Before you commit
Don't take the numbers on trust.
The engine is compiled, so we made the harness public instead. Every accuracy claim on this site can be reproduced against the binary you install.
Reproduce the benchmarks
The harness, dataset and per-question verdicts are all public. Run our LOCOMO numbers yourself against the engine you installed.
Read how we measure
Which harness, which judge, which retrieval budget, and where our numbers are weakest. We always name the budget.
Run it entirely locally
One signed binary, SQLite, no API key and no account. Nothing leaves your machine while you evaluate it.