A memory that remembers.¹

jeffs-brain/memory is a cross-language memory library for LLM agents: a shared HTTP protocol and storage spec, implemented faithfully in TypeScript, Python, and Go. Local-first, hosted-optional, Apache-2.0.

Start with TypeScript Hook up MCP See the benchmark

1

Tri-SDK smoke at 19/20 (95%); LongMemEval target 93.4% against the TypeScript baseline. Spec currently tracks v1.0 · 28/29 HTTP conformance cases green.

TypeScript

The reference implementation. CLI, library, MCP server, and Postgres / OpenFGA adapters. Drop into any Node or Bun stack.

Go

A native Go port that mirrors the reference bit-for-bit, including every retrieval regex and the full LongMemEval replay path.

III · pip / uv

Python

Idiomatic Python API across store, query, search, retrieval, and the memory pipeline. Same CLI, same wire surface.

IV

Concepts

Brains, the five memory stages (remember, extract, reflect, consolidate, recall), the retrieval ladder, and the query DSL. Start here if you want to understand how the pieces fit.

V

MCP Integration

Stdio servers for Claude Code, Claude Desktop, Cursor, Windsurf, and Zed. Twelve memory_* tools with identical wire semantics across runtimes.

CLI Reference

Every memory subcommand, flag, and exit code. Parity matrix across the three SDKs so you know what's shipped and what's stubbed.

Protocol Spec

The language-neutral contract: HTTP endpoints, storage semantics, query grammar, retrieval algorithms, MCP tool cards.

LongMemEval

The cross-SDK benchmarking harness that drives every SDK's daemon through the same question set and scores the result. 93.4% target.

Examples

Verbatim, runnable hello-worlds in TypeScript and Go. Copy, paste, hack.