noumen — coding agents are products. This is the library.

Persist knowledge across sessions with a file-backed memory provider and automatic LLM-driven extraction.

The memory system lets agents retain knowledge across sessions. Memories are stored as markdown files and optionally injected into the system prompt so the agent has access to user preferences, project conventions, and past decisions.

Configuration

import { Agent, FileMemoryProvider } from "noumen";
import { LocalSandbox, LocalFs } from "noumen/local";

const memoryFs = new LocalFs({ basePath: ".noumen/memory" });

const code = new Agent({
  provider,
  sandbox: LocalSandbox({ cwd: "/my/project" }),
  options: {
    memory: {
      provider: new FileMemoryProvider(memoryFs),
      autoExtract: true,       // run extraction after each turn
      maxIndexLines: 200,      // truncate index after 200 lines
      injectIntoSystemPrompt: true,
    },
  },
});

MemoryProvider interface

Any storage backend can be used by implementing MemoryProvider:

interface MemoryProvider {
  loadIndex(): Promise<string>;
  loadEntry(path: string): Promise<MemoryEntry | null>;
  saveEntry(entry: MemoryEntry): Promise<void>;
  removeEntry(path: string): Promise<void>;
  listEntries(): Promise<MemoryEntry[]>;
  search(query: string): Promise<MemoryEntry[]>;
}

FileMemoryProvider

The built-in FileMemoryProvider stores memories as markdown files on a VirtualFs. Each memory is a .md file with YAML frontmatter containing metadata, and a MEMORY.md index file lists all entries.

import { FileMemoryProvider } from "noumen";
import { LocalFs } from "noumen/local";

const provider = new FileMemoryProvider(
  new LocalFs({ basePath: ".noumen/memory" }),
  { maxIndexLines: 200 }
);

await provider.saveEntry({
  name: "user-preferences",
  description: "User prefers TypeScript strict mode",
  type: "user",
  content: "Always use strict TypeScript. Prefer const over let.",
});

Memory types

Type	Purpose
`user`	User preferences and personal conventions
`project`	Project-specific patterns and architecture decisions
`feedback`	Corrections and feedback from past interactions
`reference`	Documentation snippets and reference material

Auto-extraction

When autoExtract is enabled, the agent uses the LLM to analyze conversation turns and automatically extract memories worth persisting. Extraction creates, updates, or deletes memory entries based on new information.

MemoryEntry type

interface MemoryEntry {
  name: string;
  description: string;
  type: "user" | "project" | "feedback" | "reference";
  content: string;
  path?: string;
  updatedAt?: string;
}