HN Showcase
Back to browse
Engram – Open-source agent memory that beats Mem0 by 20% on LOCOMO

Engram – Open-source agent memory that beats Mem0 by 20% on LOCOMO

by tstockham·Feb 25, 2026·2 points·0 comments
Visit ProjectView on HN

AI Analysis

●●●BangerZero to OneSolve My ProblemWizardry

Read-time memory consolidation beats Mem0 on benchmark; MCP + TypeScript + SQLite, deploy anywhere.

Strengths
  • •Spreading activation cascades context automatically (query Thomas → marathon → morning runs); Mem0 extracts static facts
  • •Sleep cycles distill episodes into typed knowledge graphs instead of raw text compression—genuinely different architecture
  • •MCP integration + npm install-and-go UX means Claude/Cursor users get 10 memory tools instantly, zero boilerplate
Weaknesses
  • •80% LOCOMO vs. 66.9% Mem0 is strong, but LOCOMO is a narrow benchmark; real-world long-horizon recall untested
  • •Early adoption risk: 1.9k npm downloads, TypeScript-only, SQLite durability story underspecified for production agents
Category
AI/ML
Target Audience

AI engineers building agents, Claude/Cursor users, LLM app developers

Similar To

Mem0 · Zep · LangChain memory modules

Post Description

I built Engram because every AI agent I worked with forgot everything between sessions. Existing solutions (Mem0, Zep) are Python-first and extraction-based. They aggressively compress conversations into facts at write time.

Engram takes the opposite approach: store memories with rich metadata and invest intelligence at read time, when you actually know the query. TypeScript, SQLite, zero infrastructure.

Ran the LOCOMO benchmark (same one Mem0 used to claim SOTA):

Engram: 80.0% (10 conversations, 1,540 questions) Mem0 published: 66.9% 93.6% fewer tokens than full-context approaches

Works as an MCP server, REST API, or embedded SDK. Supports Gemini, OpenAI, Ollama, Groq, and any OpenAI-compatible provider.

npm install -g engram-sdk && engram init

https://engram.fyi | https://github.com/tstockham96/engram

Similar Projects