Feature Cascading Memory

Feature: Cascading Memory Engine (The Waterfall)

⚠️ STATUS: UPDATED (v1.41.0) The legacy 4-tier system has been decommissioned. A.I.M. now utilizes the 5-Stage Waterfall Pipeline with configurable intervals and the Consume & Clean protocol.

1. Overview

The Cascading Memory Engine is the digestive system of A.I.M. It takes the "Raw mind" (JSON terminal logs) and processes it through multiple filters until it becomes "Durable Truth" (the MEMORY.md file).

2. The 5-Stage Waterfall

The refinement process is governed by time-based triggers. Each stage only processes the "unrefined" output of the stage below it.

1. Tier 1 (Harvester): Runs every 1 hour. Extracts the "Signal Skeleton" and produces reasoning-backed ARC reports (Adds/Removes/Contradicts).
2. Tier 2 (Proposer): Runs every 12 hours. Consolidates the hourly reports into a structured Delta.
3. Tier 3 (Refiner): Runs every 24 hours. Synthesizes multiple proposals into a single Daily State.
4. Tier 4 (Consolidator): Runs every 72 hours (3 Days). Distills the daily states into macro architectural milestones.
5. Tier 5 (Archivist): Runs every 144 hours (6 Days). Converts the weekly milestones into dense, factual axioms and automatically applies them to the durable memory file.

3. The "Consume & Clean" Protocol

To prevent redundancy and token burn, each stage "consumes" its inputs.

• Once Tier 2 has successfully refined the Tier 1 reports, it immediately Archives or Deletes those reports.
• This ensures that every refinement turn is only looking at new data.

4. Manual Override (aim commit)

The operator does not have to wait for the waterfall. You can run aim commit at any time to instantly trigger an AI-driven merge of the latest refinement proposal into your durable memory.

5. Configuration

All intervals and the cleanup mode (Archive vs Delete) are live-configurable via:

aim tui  # Option 13: Configure Waterfall Pipeline

---

For the complete technical breakdown, see: Waterfall Memory Architecture