Supermemory Tops Three Authority Benchmarks, Solving AI Forgetting

Problem

When a new conversation starts, AI instantly loses previously agreed project details—technology stack, coding conventions, personal preferences—forcing developers to repeat context for every new project. Building a custom RAG memory also requires managing vector databases, embedding models, document chunking, and index tuning, often taking 3–7 days. Existing memory solutions such as LangChain and Letta involve complex configuration and lack native multimodal support.

Benchmark results (2026)

Supermemory ranks first on three authoritative memory benchmarks:

LongMemEval (long‑term memory)

LoCoMo (long conversation memory)

ConvoMo (context management)

Comparison of memory solutions

Supermemory – One‑click zero‑config, cloud or on‑premise option; no vector DB, no chunking, free web UI; first on all three benchmarks; suitable for all scenarios (personal chat, coding agents, enterprise knowledge bases); drawback: sensitive data requires the local superlocalmemory variant.

Self‑built vector DB (Pinecone/Qdrant) – Very high deployment complexity; requires selection of vector DB, embedding model, sharding strategy, similarity tuning; benchmark performance generally below averages; suited for heavy enterprise customization; drawbacks: long development cycle, high ops cost, poor multimodal compatibility.

LangChain Memory – Medium‑high complexity; manual memory chain, embedding, document split; many version‑compatibility pitfalls; medium benchmark score, long‑term memory often lost; suited for simple demos and short sessions; drawback: performance drops sharply with long context and no native multimodal extraction.

Letta Memory – Medium complexity; depends on an extra Agent framework, memory rules are cumbersome; good for specific cases but overall behind Supermemory; niche custom agents; drawback: few ecosystem connectors and high integration cost.

Unified memory architecture

Supermemory replaces the traditional three‑stage “vector DB + RAG + custom memory” stack with a unified memory layer that automatically extracts key facts (personal preferences, code choices, project architecture, troubleshooting tips) and stores them in a three‑tier hierarchy:

User Profile (global across projects) : writing style, model preferences, coding habits, common rules.

Project Memory (isolated per project) : build commands, dependency choices, bug solutions, architectural conventions (e.g., Bun replaces Node, Pinia for state).

Session Summary : concise per‑turn summary, automatic compression to avoid token overflow.

Core technologies

Automatic Learning Engine : real‑time extraction without explicit remember commands; supports both auto and manual modes.

RAG + Memory Mixed Search : a single request queries both knowledge base and personal memory, delivering ~50 ms average latency (worst‑case ≤300 ms).

Full‑Format Multimodal Extractor (2026 addition) : native parsing of PDF/PPTX (preserving outline), images (OCR), audio/video (auto transcription via Gemini 2.5 Flash), and code files (AST analysis).

All‑Ecosystem Connectors : one‑click integration with Google Drive, Gmail, Notion, OneDrive, GitHub (selective folder sync), and a web crawler.

Context Compression + Privacy Tags : automatic compression when usage exceeds 80% of the context window; <private>…</private> tags hide secrets from storage, with the local superlocalmemory version keeping data entirely offline.

Installation and configuration

No‑code web experience : visit app.supermemory.ai, register, upload documents or paste conversations, and start using the free tier.

OpenCode / Claude plugin :

# Interactive install (recommended)
 bunx opencode-supermemory@latest install

# Silent script install
 bunx opencode-supermemory@latest install --no-tui

Obtain an API key at console.supermemory.ai/keys and set it:

# Mac/Linux (zsh)
 echo 'export SUPERMEMORY_API_KEY="sm_xxx"' >> ~/.zshrc && source ~/.zshrc

# Windows (PowerShell)
 [System.Environment]::SetEnvironmentVariable("SUPERMEMORY_API_KEY","sm_xxx","User")

Optional parameters are stored in ~/.config/opencode/supermemory.jsonc (e.g., similarityThreshold, maxMemories, compactionThreshold, keywordPatterns).

MCP server standard integration (Claude Desktop / Cursor) :

docker run -p3000:3000 supermemoryai/supermemory

Browser plugin : install the official Chrome/Edge extension to auto‑save webpages, selected text, and retrieve history on demand.

Practical use cases

Developers maintaining single or multiple codebases who want AI to remember project conventions without repeated prompts.

Personal knowledge bases aggregating Notion, cloud drives, and web snippets for cross‑source AI retrieval.

Product or operations teams that need to embed work methods and requirement specifications into LLM conversations.

Batch processing of PDFs, audio/video, and images where AI must automatically summarize content.

Scenarios not recommended

Rapid prototyping with constantly changing conventions—memory churn makes benefits minimal.

Highly confidential projects—prefer the offline superlocalmemory variant to avoid any cloud storage.

One‑off tasks with no need for knowledge retention.

Developer selection guide

Free personal use : web UI + browser plugin, zero cost for daily learning and document organization.

Programming development (OpenCode / Claude) : cloud plugin with modest API fees; switch to superlocalmemory for sensitive codebases.

Enterprise private deployment : Docker/K8s deployment of the open‑source core, integrate with internal OA or knowledge bases, keep all data on‑premise.

Offline local version

The 2026‑added superlocalmemory runs entirely offline, requires no API key, and stores all memories on local disk, addressing data‑exfiltration concerns for classified projects.

Command‑line examples

# Add a project memory
supermemory add "项目使用Bun构建，禁用Node.js" --scope project --type project-config

# Search project memory
supermemory search "项目构建命令" --scope project

# View user profile
supermemory profile

# List all memories in the current project
supermemory list

# Forget a specific memory
supermemory forget "废弃的v1构建规则"

Configuration file example

{
  "apiKey": "sm_xxx",
  "similarityThreshold": 0.6,
  "maxMemories": 5,
  "maxProjectMemories": 10,
  "maxProfileItems": 5,
  "injectProfile": true,
  "compactionThreshold": 0.8,
  "keywordPatterns": ["记住", "save this", "don't forget", "note this"]
}