◆ Open-source · Privacy-first · MCP

Memory that thinks by association, not keywords

Long-term memory for Claude Desktop & Cursor, built on Kanerva's Sparse Distributed Memory — the associative model recent research proved equivalent to Transformer Attention. It recalls the meaning behind a vague prompt, even from noise.

Install free (local) ★ Star on GitHub

100% open-source · Your memories never leave your machine · $5/mo optional cloud sync

 Claude · Holographic Memory
Week 1
I really don't like Python.
Noted — I'll remember that. 🧠
Week 4 · different chat
What should I write this quick script in?
Go — you told me you dislike Python, and Go fits a quick CLI well.
↺ recalled by association · no keyword match

Plain RAG can't do that. Interference can.

Classic vector RAG is literal: no keyword overlap, no hit. SDM smears every fact across a 10,000-dimensional space and reconstructs the signal by majority vote inside the activation radius.

Vector RAGHolographic Memory (SDM)
Match modelkeyword / cosineassociative interference
Vague querymissesreconstructs from noise
Conflicting factssilently coexistflagged as interference
Foundationad-hoc embeddingsKanerva SDM ≈ Attention

Four tools your agent will actually use

Declared over MCP (stdio). Each one does something a plain database can't.

recall_by_association

Associative recall

Retrieve a de-noised "meaning cloud" from a vague or emotional cue — instantly remember context mentioned only in passing.

store_holographic_snapshot

Distributed write

Store a fact with context, emotional valence, importance and tags, superposed across the association cloud.

interference_analysis

Conflict detection

Spot when a new fact contradicts an existing belief and surface it — "You moved? I had London on file."

consolidate_and_prune

Sleep & compress

Drop weak associations, reinforce frequent ones — keeps the store fast and free of clutter.

"You moved to Berlin? I remembered London — want me to update that?"

Interference detection makes the agent feel human — it notices when your story changes.

Free to run. $5 to forget nothing, everywhere.

The engine is open-source and free forever. Pay only for convenience open code can't copy.

Local
Free
Open-source · AGPL-3.0
  • Full SDM engine + all 4 tools
  • Local SQLite store
  • 100% private — nothing leaves your machine
  • Runs in Claude Desktop & Cursor
Install
Cloud / Pro
$5/mo
Everything in Local, plus
  • Encrypted cross-device sync
  • Same memory in Claude at work & Cursor at home
  • Managed hosting + automatic backups
  • Semantic-cloud visualization
Get Pro
Business
Custom
B2B
  • Dual-licensing for closed-source products
  • Custom SDM integrations
  • Priority support & consulting
  • SIEM / logs & game-engine modules
Contact

One command to remember

Free local mode needs no key. Add a license key only for encrypted cloud sync.

terminal — via Smithery
npx -y @smithery/cli install holographic-memory
claude_desktop_config.json
{
  "mcpServers": {
    "holographic-memory": {
      "command": "uvx",
      "args": ["holographic-memory-server"],
      "env": {
        "MEMORY_MODE": "LOCAL",              // free, private
        "MEMORY_LICENSE_KEY": ""             // optional — cloud sync
      }
    }
  }
}

InsightsСтатьи

Notes on associative memory, the math behind it, and where it's going. Заметки об ассоциативной памяти, математике за ней и о том, куда всё движется.

01 · Fundamentals

Why associative memory beats vector RAG

Vector RAG matches on surface similarity: no shared keywords, no hit. It cannot connect "I dislike Python" to a later "what should I write this script in?" — the words don't overlap.

Holographic memory smears each fact across a 10,000-dimensional space and reconstructs the signal by majority vote inside the activation radius. It survives noise and vague cues, and surfaces the connection RAG never sees: "Go — you dislike Python."

Почему ассоциативная память сильнее векторного RAG

Векторный RAG ищет по поверхностному сходству: нет общих слов — нет результата. Он не свяжет «не люблю Python» с более поздним «на чём написать скрипт?» — слова не пересекаются.

Голографическая память «размазывает» факт по 10 000-мерному пространству и восстанавливает сигнал по правилу большинства внутри радиуса активации. Она устойчива к шуму и смутным запросам и находит связь, которую RAG не видит: «Go — ты же не любишь Python».

02 · Theory

Sparse Distributed Memory ≈ Attention

The engine is Kanerva's SDM, first described in MIT Press. Recent work (2021–2026) proved it is mathematically equivalent to the Attention mechanism inside Transformers — the very thing that powers GPT-4 and Claude.

So this isn't exotica bolted onto an LLM. It's the same principle, exposed as durable, write-once-recall-forever memory the model can reach through MCP.

Разреженная распределённая память ≈ Attention

В основе — SDM Канервы, впервые описанная в MIT Press. Недавние работы (2021–2026) доказали её математическую эквивалентность механизму Attention в трансформерах — тому самому, на котором работают GPT-4 и Claude.

Это не экзотика сбоку от LLM, а тот же принцип — только вынесенный в долговременную память, к которой модель обращается через MCP.

03 · Frontier

From AI memory to a robot's "muscle memory"

Kanerva designed SDM as a digital model of the cerebellum — the seat of motor skill. For narrow, motor tasks (walking, skating) that turns a weakness into a strength: repetition builds a stable interference pattern. The more it walks, the better it walks.

The far horizon: bake a trained skill into a physical optical hologram and read it at the speed of light — a "crystal of parkour" for humanoid robots and drones. Today: simulation. But the vector is set.

От памяти ИИ к «мышечной памяти» робота

Канерва задумывал SDM как цифровую модель мозжечка — центра моторики. Для узких моторных задач (ходьба, катание) это превращает минус в плюс: повторение формирует устойчивую интерференционную картину. Чем больше ходишь — тем лучше ходишь.

Дальний горизонт: записать навык в физическую оптическую голограмму и считывать его со скоростью света — «кристалл паркура» для человекоподобных роботов и дронов. Пока — симуляция. Но вектор задан.

More in the full project story → Подробнее — в полной истории проекта →

Partnership & demoПартнёрство и демо

Want a live demo, a commercial license, or to build something together? Send a note. Нужна демонстрация, коммерческая лицензия или совместный проект? Напишите нам.

Opens your email app addressed to the team. Prefer GitHub? Open an issue. Откроется ваш почтовый клиент с письмом команде. Удобнее через GitHub? Создайте issue.