EngramEngramdocs
v0.1.0
Search docs…⌘K
GitHub
v0.1.3 — Open Source, MIT

Engram Documentation

Engram is a persistent memory layer for any AI model. It stores episodic, semantic, and procedural memories in a local SQLite database, retrieves them with sub-100 ms latency using vector search, and injects them into AI context automatically via MCP or HTTP.

10 k+
writes / sec
< 100 ms
recall latency
384-dim
local embeddings
0
API keys required

What is Engram?

Modern AI assistants forget everything the moment a conversation ends. Engram solves this by maintaining a universal brain — a single, growing knowledge store shared across every tool you use.

Any AI that speaks MCP (Model Context Protocol) or HTTP can connect to Engram. Once connected, memories are automatically stored and retrieved on every interaction — no prompt engineering, no manual copy-paste, no cloud required.

Architecture

Engram runs as a local daemon (port 4901 by default). It exposes two surfaces:

  • MCP server — MCP server — consumed by Claude Code, Claude Desktop, Cursor, and any MCP-compatible client
  • REST API — REST API — consumed by custom apps, Ollama proxy, and direct HTTP integrations
AI client ──MCP──▶ engram daemon ──▶ SQLite + HNSW REST client ──HTTP──▶ engram daemon ──▶ vector search Ollama client ──proxy──▶ :11435 proxy ──▶ context inject

Core Concepts

Episodic
Timestamped events — conversations, decisions, observations. Who said what, when.
Semantic
Factual knowledge — concepts, preferences, relationships. What you know to be true.
Procedural
Trigger → action patterns. When X happens, do Y. Learned workflows and habits.
NextInstallation →
Introduction