🧠 AntiMem — Edge-Native Memory Daemon & CRDT Compactor

An industrial-grade background memory daemon and token compactor. Silently index, monitor, and inject semantic insights into active context windows via ChromaDB and FileSystem Watchdogs.

Quickstart • The Paradigm Shift • CLI Reference • Architecture • Subsystems • API Reference • Comparison Matrix • Roadmap

---

💡 The Paradigm Shift

AI agents fail when their conversational context grows too large, causing LLM context limits to trigger or query latency to spike. Standard memory retrieval engines are too slow, relying on heavy REST network requests that interrupt active coding execution loops.

AntiMem resolves this by shifting memory management to a lightweight local edge daemon:

1. Real-time Event Watchers: Uses OS-level file system notifications to monitor active conversation transcripts in the background without poll loops.
2. Local Vector Caching: Stores, index, and retrieves semantic matches instantly from an embedded local ChromaDB instance using fast inference.
3. Canary-Gated Compaction: Employs an 80/20 CRDT tombstone compression pipeline that reduces log sizes by 80% while retaining critical entity structures.
4. Cross-Platform Lock Guards: Utilizes platform-agnostic file lock mechanisms (msvcrt on Windows, fcntl on Unix) to prevent concurrency race conditions.

---

⚡ Quickstart

Prerequisites

• Python 3.11+
• uv (recommended for rapid dependency syncing)
• Ollama running locally on your system for LLM summarization

1. Clone & Setup

git clone https://github.com/axtontc/AntiMem.git
cd AntiMem

# Sync virtual environment using uv
uv sync

2. Verify with the Test Suite

Ensure the indexer, prompt extraction rules, and compactor work locally:

uv run python -m pytest tests/ -v

---

🛠️ CLI Reference

1. Start the Memory Daemon

Launch the background watcher daemon to scan skills and actively monitor conversation transcripts:

antimem start

2. Compact Context File

Compresses logs exceeding token boundaries by replacing the oldest 80% with an LLM-synthesized summary (CRDT Tombstone):

antimem compact --file "C:/path/to/transcript.jsonl" --threshold 4000

---

🏗 Architecture

flowchart TD
    A[Agent Workspace File] -->|OS File System Event| B[Watchdog Event Handler]
    B -->|Submit Task| Pool[ThreadPoolExecutor]
    Pool -->|Extract Prompts| C[ChromaDB Indexer]
    C -->|Retrieve Past Insights| D[Semantic Query]
    D -->|Ollama Synthesis Prompt| E[LLM Local Synth]
    E -->|JSON Payload| F[JSON File Injection]
    F -->|Write-Ahead File Lock| A

    style B fill:#1a1a2e,stroke:#3776AB,color:#fff
    style Pool fill:#16213e,stroke:#3776AB,color:#fff
    style C fill:#16213e,stroke:#3776AB,color:#fff
    style F fill:#2ea043,stroke:#2ea043,color:#fff

Loading

---

🏗️ Core Subsystems

Subsystem	Folder / File	Responsibility
CLI Router	antimem/cli.py	Routes arguments to starting watchers or triggering compactors
Watcher Daemon	antimem/core/daemon.py	Initializes persistent ChromaDB collections, parses directories, and manages watchdog events
CRDT Compactor	antimem/core/compactor.py	Performs 80/20 file splits, generates LLM summaries, and writes locked tombstone divisions
Skill Registry	antimem/core/daemon.py	Scans workspace markdown configurations to automatically extract and register capabilities

---

📖 API & Core Functions Reference

antimem/core/daemon.py

These functions manage background watchers and semantic indexing:

Function / Method	Parameters	Description
AntimemIndexer.index_skills()	None	Discovers and indexes markdown headers and description profiles in global config spaces.
AntimemIndexer.query(text, limit)	str, int	Conducts query matching on vector spaces returning high-similarity results.
TranscriptHandler.on_modified(event)	FileSystemEvent	Watchdog routine detecting modification logs, extracting prompts, and injecting synth context.

antimem/core/compactor.py

This class drives file context compaction:

Function / Method	Parameters	Description
Compactor.compact()	None	Locks target files, runs token heuristics, prompts Ollama, and commits tombstones.

---

📊 Comparison Matrix

Feature	Standard Chroma	pgvector-only	memmcp	AntiMem
Edge Watchdog Triggers	❌	❌	❌	✅ Yes (Watchdog)
80/20 CRDT Compactor	❌	❌	❌	✅ Yes (Tombstones)
Background Thread Injection	❌	❌	❌	✅ Yes (Non-blocking)
Fast Local Storage	✅ Yes	❌	✅ Yes	⚡ High performance (<3ms local)

---

🧰 Tech Stack

• Core Language: Python 3.11+
• Vector Store: ChromaDB (persistent client)
• File Monitor: watchdog (OS-level notifications)
• Model Inference: Ollama (Qwen2.5-coder & nomic-embed-text)
• Testing & Formatting: pytest, Ruff, mypy

---

🗺️ Roadmap

• Persistent local ChromaDB integration
• OS-level Watchdog event listening loop
• Non-blocking thread pool execution
• 80/20 CRDT tombstone log compression
• Cross-platform file locking support
• Increment Vector Deletes — Implement direct deletion pipelines when files are modified or deleted
• Smarter Tokenizer — Switch from character heuristics to tiktoken embeddings calculations
• Visual Admin UI — Add a micro-dashboard to inspect indexed skills and memories visually

---

🔗 Ecosystem Cross-Linking

AntiMem acts as the context injection daemon of the Antigravity Swarm ecosystem:

Project	Description
AUI	Zero-latency cross-process UI automation for Windows and Web
MemMCP	Stdio-transport memory server using SQLite WAL and FAISS RRF
The-Skillbrary	FastMCP skill execution server and registry
Multiverse-Planner	Brute-forces optimal plans via timeline expansion and pruning
Fractal-Swarm-v2	Mathematically optimal state-machine agent swarm orchestration
OmniMem	PostgreSQL hybrid memory system for large enterprise swarms
The-Sentinel-Reviewer	Code quality gatekeeper running static audits, tests, and deep traces

---

📜 License & Copyright

This project is licensed under the Apache License, Version 2.0. See the LICENSE file for details. Copyright (c) 2026 Axton Carroll.

---

⭐ If AntiMem helps keep your agentic contexts clean, consider giving it a star!



_{Built by Axton Carroll — "Nothing is impossible, we merely don't know how to do it yet."}