Skip to main content
Glama
holographic-memory

holographic-memory

by neo37

🧠 Holographic Memory — MCP Server (Go)

🌐 Website: holo.ai3d.art  ·  Live (GitHub Pages): neo37.github.io/holographic-memory  ·  Open-source · Privacy-first · $5/mo cloud

The first fully open-source, privacy-first holographic long-term memory for AI agents. Built on Kanerva's Sparse Distributed Memory (SDM) — the associative memory that recent research (2021–2026) proved to be mathematically equivalent to the Attention mechanism inside Transformers (GPT-4, Claude).

Give Claude Desktop, Cursor and any MCP-compatible agent a memory that thinks by association, not by keyword match. Say "I don't like Python" today, ask "what should I write this script in?" next month — and the agent recalls "Go, because you don't like Python." Plain vector RAG can't do that. Interference-based recall can.


Table of Contents / Оглавление

#

English

Русский

1

Why Holographic Memory

Зачем голографическая память

2

How It Works

Как это работает

3

The Math

Математическая модель

4

MCP Tools

Инструменты MCP

5

Architecture

Архитектура

6

Editions & Pricing

Редакции и цены

7

Install

Установка

8

Tech Stack

Технологический стек

9

Roadmap

Дорожная карта

10

Documentation

Документация

11

License

Лицензия


Related MCP server: Memsolus MCP Server

1. Why Holographic Memory

Classic RAG is literal: no keyword overlap → no hit. SDM stores every fact as a high-dimensional binary vector ({0,1}ⁿ, n ≈ 10 000) smeared across many addresses. Recall reconstructs the signal by majority vote over everything inside the activation radius, so it survives noise, partial cues and vague prompts — and it surfaces connections the user only hinted at.

Vector RAG

Holographic Memory (SDM)

Match model

keyword / cosine similarity

associative interference

Vague query

misses

reconstructs from noise

Conflicting facts

silently coexist

flagged as interference

Foundation

ad-hoc embeddings

Kanerva SDM ≈ Transformer Attention

2. How It Works

flowchart LR
    A["Fact:<br/>'User dislikes Python'"] -->|encode| B["Hypervector<br/>{0,1}^10000"]
    B -->|"write into radius r"| C[(Distributed<br/>address cloud)]
    Q["Vague query:<br/>'what language?'"] -->|encode| D["Query vector"]
    D -->|"activate within r"| C
    C -->|"majority-rule read"| E["De-noised recall:<br/>'Use Go — you dislike Python'"]

A fact is not stored in one row — it is superposed across every hard location within a Hamming radius. Reading a noisy or vague cue re-collects those overlapping traces and votes them back into a clean answer.

3. The Math

Implemented in Go, straight from Kanerva's SDM:

  • Distance — Hamming: &nbsp; d(A, B) = Σᵢ (Aᵢ ⊕ Bᵢ)

  • Write — interference: activate every hard location within radius r of address X, then increment/decrement their counters (wave superposition): Activate(X) = { Y ∈ HardLocations | d(X, Y) ≤ r }

  • Read — associative recall: sum activated cells around query Q, apply the majority rule: Outputᵢ = sign( Σ_{Y ∈ Activate(Q)} CellContents(Y)ᵢ )

This reconstructs a 100%-clean context even from a noisy or partially forgotten query.

4. MCP Tools

Tool

What it does

store_holographic_snapshot

Store a structured memory (fact + context + emotional valence + importance + tags) as a superposed hypervector.

recall_by_association

Retrieve a de-noised "meaning cloud" from a vague or emotional cue.

interference_analysis

Detect when a new fact collides with an existing belief; return the conflict + confidence.

consolidate_and_prune

"Sleep": drop weak associations, reinforce frequently used ones, keep the store fast.

{
  "name": "recall_by_association",
  "arguments": { "query": "the project I worked on when I felt down", "association_depth": 3 }
}
{ "name": "interference_analysis", "arguments": { "new_fact": "I moved to Berlin" } }
// → { "conflict_detected": true, "previous_memory": "User lives in London", "confidence": 0.85 }

5. Architecture

flowchart TB
    subgraph Client["AI Agent — Claude Desktop / Cursor"]
        AG[LLM Agent]
    end
    subgraph Server["Holographic Memory Server (Go)"]
        MCP["MCP handler<br/>(stdio / JSON-RPC)"]
        LIC{"License gate<br/>LOCAL = free"}
        SDM["SDM Engine<br/>encode · write · recall"]
        STORE[("SQLite / binary<br/>association store")]
    end
    CLOUD["☁️ Cloud Sync (Pro $5/mo)<br/>encrypted cross-device"]

    AG <-->|"tools/call"| MCP
    MCP --> LIC --> SDM --> STORE
    SDM -. optional .-> CLOUD

6. Editions & Pricing

This project ships Open-Core: the engine is free and open, convenience is paid.

flowchart LR
    Free["🆓 Local — Free<br/>MIT/Apache-2.0<br/>Full SDM engine · 4 tools<br/>Local SQLite · 100% private"]
    Pro["⭐ Cloud / Pro — $5/mo<br/>Encrypted cross-device sync<br/>Managed hosting + backups<br/>Semantic-cloud viz"]
    Biz["🏢 Business<br/>Dual-licensing<br/>Custom SDM integrations"]
    Free --> Pro --> Biz
  • Local (Free) — runs 100% on your machine; your memories never leave your computer.

  • Cloud / Pro ($5/mo) — same memory in Claude at work and Cursor at home; managed, backed up, and visualized.

  • Business — closed-source embedding rights + bespoke integrations.

7. Install

# One command via Smithery
npx -y @smithery/cli install holographic-memory

Or add it manually to claude_desktop_config.json:

{
  "mcpServers": {
    "holographic-memory": {
      "command": "uvx",
      "args": ["holographic-memory-server"],
      "env": {
        "MEMORY_MODE": "LOCAL",
        "MEMORY_LICENSE_KEY": "optional — only for Cloud/Pro sync"
      }
    }
  }
}

MEMORY_MODE=LOCAL needs no key and is free forever. Set a MEMORY_LICENSE_KEY (get one at holo.ai3d.art) only to unlock encrypted cross-device sync.

8. Tech Stack

  • Go 1.24+ — fast, low RAM, single static binary

  • MCP over stdio (JSON-RPC)

  • Local storage — SQLite / binary association file implementing Kanerva SDM

  • Docker — multi-stage Alpine build

  • Payments — Lemon Squeezy (license keys + subscriptions)

9. Roadmap

Tier

Focus

Status

1

Long-term memory for Claude Desktop / Cursor

🚧 In progress

2

Game engines (Unity / Unreal) — NPC skeletal "muscle memory"

🔭 Planned

B2B

Logs / SIEM anomaly detection (patterns smeared across time)

🔭 Planned

Full timeline & Gantt: see docs/GTM_PLAN.md.

10. Documentation

11. License

Dual-licensed:

  • AGPL-3.0 (free) — personal, self-hosted, and open-source use. If you run a modified version as a network service, AGPL requires you to publish your corresponding source. See LICENSE.

  • Commercial License (paid) — required to embed this software in a closed-source or commercial product, or to run it inside a proprietary service without publishing your source. Get it at holo.ai3d.art. See COMMERCIAL-LICENSE.md.


"Smart long-term memory for Claude that doesn't forget the context of a chat from a week ago."

Available Tools

4 tools
consolidate_and_pruneC

Sleep: drop weak/unused associations, keep the store fast.

Inputs: max_loss_tolerance.

interference_analysisB

Check whether a new fact collides with an existing belief; returns the conflict and confidence.

Inputs: new_fact (required).

recall_by_associationB

Retrieve a de-noised meaning cloud from a vague or emotional cue. Use when the user refers to the past indirectly.

Inputs: query (required), top_k, association_depth.

store_holographic_snapshotB

Store a structured memory (fact + context + emotional valence + importance + tags) as a superposed hypervector.

Inputs: fact (required), tags, context, valence, +1 more.

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    TDQS

    B3.4/5.0

    Scored across 4 tools

    Disambiguation5/5

    Each tool has a distinct purpose: storing, recalling, analyzing conflicts, and pruning. No overlap in functionality.

    Naming Consistency5/5

    All tool names follow a consistent verb_noun pattern with underscores, making them predictable and readable.

    Tool Count5/5

    Four tools cover the core operations of a memory server without being too few or excessive.

    Completeness4/5

    The tools cover storage, retrieval, conflict analysis, and maintenance, but lack explicit delete or update operations, leaving minor gaps.

    Maintenance

    ActivityInactive
    ResponsivenessOutdated data

    Related MCP Connectors

    Related MCP Servers

    • F
      license
      Not graded
      quality
      D
      maintenance
      Provides AI agents with persistent, searchable memory that survives across conversations using semantic search, temporal versioning, and smart organization. Enables long-term context retention and cross-session continuity for AI assistants.
      14
      -
    • A
      license
      A
      quality
      D
      maintenance
      Provides persistent long-term memory for AI agents through semantic search and automated knowledge graph extraction. It enables agents to store, recall, and reason over facts, preferences, and relationships across multiple conversations and sessions.
      14
      6 npm
      MIT
    • A
      license
      Not graded
      quality
      C
      maintenance
      Enables AI assistants to have persistent associative memory across sessions by storing memories, discovering associations, and retrieving them via spreading activation.
      Apache 2.0