MCP 服务器

Engram

Perstistant Memory Layer for Agents

README

🧠 Engram

The intelligence layer for AI agents

Every AI agent is born smart but amnesiac. Engram fixes that. It doesn't just store memories -- it learns, consolidates patterns, detects contradictions, and surfaces context you didn't ask for.

Install

npm install -g engram-sdk
engram init

That's it. Works with Claude Code, Cursor, or any MCP client. Also available as a REST API and TypeScript SDK.

Why Engram

Existing memory solutions are storage layers -- they save facts and retrieve them. Engram is an intelligence layer with three tiers:

Tier	What it does	Who has it
Explicit Memory	Stores facts, preferences, conversation turns	Everyone
Implicit Memory	Detects behavioral patterns from how users work	Engram only
Synthesized Memory	Consolidation produces insights nobody asked for	Engram only

Key insight: Engram invests intelligence at read time (when the query is known), not write time (when you don't know what'll matter). This is the fundamental architectural difference from Mem0, Zep, and LangMem.

Benchmarks

Evaluated on LOCOMO -- the standard benchmark for agent memory systems. Same benchmark Mem0 used to claim state of the art.

System	Accuracy	Tokens/Query
Engram	80.0%	1,504
Full Context	88.4%	23,423
Mem0 (published)	66.9%	--
MEMORY.md	28.8%	--

10 conversations, 1,540 questions, 4 categories. 19.6% relative improvement over Mem0 with 93.6% fewer tokens than full context.

Full context (dumping entire conversation history) scores highest but uses 30x more tokens and can't scale past context window limits. Engram closes most of the gap while using 96.6% fewer tokens.

Full benchmark methodology and per-category breakdown

Features

MCP Server -- 10 memory tools for Claude Code, Cursor, and any MCP client
REST API -- Full HTTP API for any language or framework
TypeScript SDK -- Embedded use for Node.js agents
CLI -- Interactive REPL, bulk operations, eval tools
Model-agnostic -- Works with Gemini, OpenAI, Ollama, Groq, Cerebras (any OpenAI-compatible provider)
Zero infrastructure -- SQLite, no Docker, no Neo4j, no Redis
Consolidation -- LLM-powered memory merging, contradiction detection, pattern discovery
Entity-aware recall -- Knows "Sarah" in the query should boost memories about Sarah
Bi-temporal model -- Tracks when facts were true, not just when they were stored
Spreading activation -- Graph-based context surfacing

Quick Start

MCP Setup (Claude Code / Cursor)

npm install -g engram-sdk
engram init

REST API

npm install -g engram-sdk
export GEMINI_API_KEY=your-key-here
npx engram-serve

Server starts on http://127.0.0.1:3800.

Remember and Recall

# Store a memory
curl -X POST http://localhost:3800/v1/memories \
  -H "Content-Type: application/json" \
  -d '{"content": "User prefers TypeScript over JavaScript", "type": "semantic"}'

# Recall relevant memories
curl "http://localhost:3800/v1/memories/recall?context=language+preferences&limit=5"

TypeScript SDK

import { Vault } from 'engram-sdk';

const vault = new Vault({ owner: 'my-agent' });

await vault.remember('User prefers TypeScript');
const memories = await vault.recall('language preferences');
await vault.consolidate();

API Reference

Full REST API and MCP tool documentation: engram.fyi/docs

Configuration

Variable	Description	Default
`GEMINI_API_KEY`	Gemini API key for embeddings and consolidation	--
`ENGRAM_LLM_BASE_URL`	Custom API base URL (Groq, Cerebras, Ollama, etc.)	provider default
`ENGRAM_LLM_MODEL`	LLM model name	provider default
`ENGRAM_DB_PATH`	SQLite database path	`~/.engram/default.db`
`PORT`	Server port	`3800`
`ENGRAM_AUTH_TOKEN`	Bearer token for API auth	--