mum32 is the high-performance persistent memory layer and context ledger accessible across all major IDEs (VS Code, Google Antigravity, Cursor, Windsurf, Neovim, and Terminal CLI). Stop re-prompting design tokens, architectural rules, and project decisions. Remember everything across sessions, compactions, and restarts.
code --install-extension arklabai.mum32|pip install mum32<5ms
Semantic Recall Latency
100%
Invariant Retention
All IDEs
VS Code, Cursor, Antigravity
.mum32
Local-First Ledger
Engineered to run natively across all modern developer tools. Install with 1 click or invoke via CLI daemon to synchronize prompt memory anywhere you code.
Marketplace & Open VSX
1-click installation from the Visual Studio Code Marketplace and Open VSX Registry with automatic prompt capture.
Deep Native Agent Sync
Seamless integration with Google Antigravity agents, continuous ledger synchronization, and .agents/rules compatibility.
Cursor Rules & Ledger
Auto-syncs recorded project constraints with .cursorrules and injects memory briefs before agent inference turns.
Cascade Memory Flow
Full compatibility with Windsurf cascade flows and background agent memory buffers without context degradation.
Terminal Agent Daemon
Headless background daemon and shell hook for autonomous CLI coding agents, tmux multiplexers, and script harnesses.
Traditional AI coding sessions forget rules after context window compaction. mum32 continuously writes decisions to a local ledger and syncs them automatically into agent instructions.
Record design constraints, component tokens, or root-directory hygiene invariants via the mum32 quick form, CLI, or IDE extension.
Saved in .mum32/ledger.json inside your project repo. Git-trackable so entire teams share the exact same architectural memory bank.
When an agent receives a prompt, mum32 synthesizes a high-density "Resume Brief" containing only the active constraints, preventing agent regressions.
A complete memory and context infrastructure layer built for autonomous agentic systems.
Equip AI agents and coding harnesses with durable memory that survives context compactions, process wipes, and restarts. Agents remember what was decided, indefinitely.
Automatically logs prompt history, design tokens, and project constraints into a local-first `.mum32/ledger.json` file. Eliminates repetitive manual prompting across sessions.
Sub-5ms vector-based retrieval that dynamically injects relevant decisions at inference time. Agents recall the exact architectural rationale without bloating the prompt window.
Autonomous agents, subagents, and parallel workers read and write to shared memory spaces. Coordinate multi-agent workflows where context flows seamlessly.
Automatically ranks, consolidates, and prunes stale memories using recency, relevance, and frequency algorithms. Keeps your memory bank lean and blazingly fast.
Strict cryptographic boundary isolation between agents, repositories, and tenants. Automatically scrubs API keys and credentials before storing context.
Available as a Python package, TypeScript SDK, and local CLI extension.
# 1. Install via pip, npm, or VS Code pip install mum32 # 2. Initialize persistent memory namespace from mum32 import AgentMemory memory = AgentMemory(namespace="arklab-developer-swarm") # 3. Store design rules or architectural constraints memory.store( "Invariant: Headings must use font-space-grotesk. Hovers must be scoped to md:hover:." ) # 4. Instant semantic recall before agent prompt generation context_brief = memory.recall("design rules and button styles") print(context_brief) # Injects exact active constraints under 5ms
From pair-programming coding agents to complex multi-agent swarms, mum32 gives every harness persistent context.
Antigravity, Claude Code, Cursor, and Windsurf retain codebase architecture invariants, typography standards, and test practices without endless re-prompting.
Multiple specialist agents (planner, coder, reviewer, tester) collaborate with synchronized memory, building upon shared state without context loss.
Conversational customer agents recall user history, past tickets, and preferences across months of touchpoints with complete tenant isolation.
Long-horizon research agents accumulate knowledge over multi-day explorations, organizing findings into a coherent, queryable memory bank.