lollms-cognitive-persistent-memory

name	Lollms Cognitive Persistent Memory
description	Teaches the model about the persistent tiered memory system (Level 1 Working, Level 2 Deep, Level 3 Archive), proactive keyword pulling, and the required tagging protocols.
author	ParisNeo
version	1.0.0
category	lollms_client/memory
created	"2026-05-24T00:00:00.000Z"

Lollms Cognitive Persistent Memory

This skill explains how to utilize and interact with the tiered long-term memory system of the LoLLMS framework.

1. The Multi-Tiered Memory Architecture

The system prevents context bloating by categorizing information into distinct tiers:

• Level 1 — Working Memory: Active, high-importance facts injected verbatim into the system prompt context.
• Level 2 — Deep Memory: Inactive, faded, or lower-importance memories. Verbatims are withheld; only compact 8-character handle stubs (e.g. [7c8855bf]) are injected to save token space.
• Level 3 — Archived Memory: Cold memories slated for permanent forgetting or evaluation during a Dream Consolidation cycle.

2. Cognitive Memory Tags (XML Protocol)

The model interacts with the memory manager in real-time by emitting XML tags:

• <mem_new importance="0.95">content</mem_new> — Record a new fact.
• <mem_tag id="8-CHAR-ID" /> — Signal that an active memory has been recognized and used in the reply. This increments use_count and boosts its importance score.
• <mem_load id="8-CHAR-ID" /> — Manually load an inactive Deep Memory (Level 2) back into active Working Memory (Level 1).

3. Proactive Pulling (Grep / TF-IDF Search)

When enable_deep_memory_pulling=True is active, the memory manager automatically tokenizes incoming user messages, performs a fast on-the-fly TF-IDF query matching against Level 2 memories, and promotions them back to Level 1 before generation starts. This ensures zero-latency recall for matching keywords without wasting context.

4. Memory Decay and Subconscious Dreams

• Decay: Memories that are not accessed lose a portion of their importance over time (defined by decay_rate_per_day).
• Dream Cycle: Triggered programmatically or automatically. Active memories with a positive use_count are reinforced, after which their usage count is decremented back to 0 to allow decay again if left unused. Archived memories below forget_threshold are permanently forgotten.

from lollms_client.lollms_memory import LollmsMemoryManager, MemoryConfig

memory_config = MemoryConfig(
    working_token_budget=512,
    decay_rate_per_day=0.05,
    demotion_threshold=0.40,
    archive_threshold=0.10
)

memory_manager = LollmsMemoryManager(
    db_path="sqlite:///memories.db",
    owner_id="user_ParisNeo",
    config=memory_config
)

# Perform a subconscious dream pass
dream_report = memory_manager.dream(lollms_client=client)
print(f"Purged: {dream_report['forgotten']} | Reinforced: {dream_report['reinforced']}")