Feature Quarantine Daemon

Phase 45: The Quarantine Daemon (Non-Destructive Context Pruning)

⚠️ STATUS: CONCEPTUAL ARCHITECTURE (BLEEDING EDGE) This document outlines a highly advanced context-management architecture designed to act as a "Non-Destructive Garbage Collector" for LLM memory arrays. We believe this pattern sits at the absolute cutting edge of agentic systems, and we are open-sourcing it for free.

1. The Market Gap

As of early 2026, managing "Context Rot" is the primary bottleneck for autonomous agents.

• The Industry Solution: Commercial tools like Claude Code and Roo Code rely on manual user intervention. They provide a /rewind command, forcing the human to realize they are in a rabbit hole and manually delete the conversation history. Once deleted, the data is gone forever.
• The Academic Solution: Research frameworks like TC-RAG use "State Pops" to rewind context when the AI detects its own logical contradictions.

The Missing Link: No framework currently handles User-Side Thrashing (human flip-flopping) autonomously while maintaining Forensic Data Safety.

A.I.M. solves this via the Quarantine Daemon.

2. The Problem: User-Side Thrashing

When an operator changes their mind during a build, it devastates the LLM's context window.

The "Flip-Flop" Scenario:

1. User: "Add a red button."
2. AI: Writes code for a red button.
3. User: "Actually, make it blue."
4. AI: Rewrites code for a blue button.
5. User: "Nevermind, just delete the button completely."
6. AI: Deletes the button code.

In a standard agentic CLI, the active context window is now poisoned with 6 turns of useless, contradictory logic and obsolete code blocks. The LLM continues to pay massive API token costs to "remember" a button that no longer exists, increasing the risk of future hallucinations.

3. The Architecture: "Ctrl+Z for AI Context"

The Quarantine Daemon operates as a background Python hook that intercepts the raw JSON chat array and performs a Soft Delete.

Step 1: Heuristic Detection

The daemon monitors the user's prompts for "Reversion Semantics" (e.g., "revert that", "nevermind", "actually go back to"). When detected, it identifies that a logical side-quest has just been invalidated.

Step 2: The Surgical Excision

Instead of forcing the LLM to write a response acknowledging the reversion, the daemon programmatically steps in. It calculates the exact conversational turn where the side-quest began, and physically slices those JSON blocks out of the active terminal session's memory array.

Step 3: The Quarantine (Soft Delete)

Silent deletion is dangerous. If the user later decides, "Wait, what was the hex code for that blue button we tried earlier?", the data must be retrievable.

The daemon does not delete the excised JSON. Instead, it appends the raw JSON blocks to a local graveyard file: continuity/private/PRUNED_GRAVEYARD.jsonl.

Step 4: Synthetic Injection

To prevent the human operator or the LLM from getting confused by the sudden jump in conversation, the daemon replaces the excised turns with a single, ultra-compressed synthetic marker:

[SYSTEM NOTE: The previous 6 turns regarding the button UI were quarantined by the daemon to prevent context bloat. To restore this data, run 'aim unprune'.]

4. The Result

By implementing the Quarantine Daemon, A.I.M. achieves two previously mutually exclusive goals:

1. Zero Context Weight: The AI's active token payload drops instantly. It stops hallucinating based on dead-end code blocks.
2. 100% Forensic Fidelity: The operator loses absolutely nothing. If an old idea becomes relevant again, the aim unprune command instantly stitches the quarantined JSON back into the active array.

It is a perfectly safe, automated garbage collection system for human logic.