najia-afk/aria_skills/memory_compression

---
name: aria-memory-compression
description: "🗜️ 3-tier hierarchical memory compression"
metadata: {"aria": {"emoji": "🗜️"}}
---

# aria-memory-compression

3-tier hierarchical memory compression engine. Scores memory importance,
compresses via LLM summarization (with rule-based fallback), and manages
raw → recent → archive tiers. Stores compressed summaries in semantic memory.

## Architecture

Raw memories (limit: 20) ↓ ImportanceScorer (recency × significance × category × length) ↓ MemoryCompressor (LLM via LiteLLM + fallback) Recent tier (limit: 100, ratio: 0.3) ↓ Archive tier (all older, ratio: 0.1) ↓ SemanticMemory (category: compressed_recent / compressed_archive)

## Usage

```bash
# Compress a batch of memories through the pipeline
exec python3 /app/skills/run_skill.py memory_compression compress_memories '{"memories": [{"content": "...", "category": "task", "timestamp": "2026-02-16T10:00:00Z"}]}'

# Compress recent session (last N hours)
exec python3 /app/skills/run_skill.py memory_compression compress_session '{"hours_back": 6}'

# Get working context within token budget
exec python3 /app/skills/run_skill.py memory_compression get_context_budget '{"max_tokens": 2000}'

# Check compression statistics
exec python3 /app/skills/run_skill.py memory_compression get_compression_stats '{}'

Functions

compress_memories

Compress a list of memories through the 3-tier pipeline. Scores importance, groups by tier, LLM-summarizes each group, stores in semantic memory. Returns compression ratio, tokens saved, and summaries.

compress_session

Quick compress of recent session activity via api_client.summarize_session(). Useful for end-of-session cleanup.

get_context_budget

Retrieve working memory context within a token budget. Includes both raw working memory items and compressed summaries from previous runs.

get_compression_stats

Get statistics from the last compression run — memories processed, compression ratio, tokens saved, tier breakdown.

Dependencies

• api_client (semantic memory storage, working memory, session summarization)
• LiteLLM proxy (kimi model for LLM summarization)