AAAK Compression

Adaptive Associative Abstraction Kernel (AAAK) is Mnemosyne's text substitution system for reducing memory content size. It is not a method on the Memory class — it is a standalone utility used during consolidation.

The Problem

Agent context windows are limited (4K–128K tokens). Raw memory dumps consume precious space. AAAK applies deterministic text substitutions to compress related memories into shorter representations.

How AAAK Works

The encode() function performs text substitution in three passes:

1. Category prefixes: Replaces verbose category labels with short tokens (e.g., "conversation:" → "conv:")
2. Phrase compression: Substitutes common phrases with abbreviated forms
3. Structural: Normalizes whitespace, removes redundant formatting

Compression Ratios

AAAK achieves approximately 30–50% text reduction:

Actual ratios vary by content. AAAK is lossy — some detail is sacrificed for brevity.

Using AAAK

from mnemosyne.aaak import encode

# Encode (compress) text
compressed = encode(original_text)

# The result is a shorter string, not a separate object
print(f"Original: {len(original_text)} chars")
print(f"Compressed: {len(compressed)} chars")

During Consolidation

AAAK is applied automatically during sleep consolidation when summarizing groups of Working Memory entries:

flowchart LR
  WM[Working Memory Entries] -->|group by source| G[Grouped]
  G -->|encode via AAAK| C[Compressed Summary]
  C -->|promote| EM[Episodic Memory]

Important Notes

• encode() is one-way: There is no decode() or expand() function. Once compressed, the original text is not recoverable from the output alone.
• Not a Memory method: AAAK is a standalone module, not called via mem.compress() or mem.expand().
• Used during consolidation: Applied by the sleep consolidation pipeline to compress summaries before promoting to Episodic Memory.