Menu
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.
Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.
Based on SOC occupation classification
Teaches how to build, discover, and execute local, zero-configuration Python tools in Lollms using AST schema parsing and context injection. Includes step-by-step tool file construction guide and code examples.
Teaches the model how to formulate strategic <coding_plan> blocks, apply indentation-agnostic Aider search/replace patches, manage Git-like version logs, and control in-message status reporting.
Teaches the model how to manage stateful discussions, track conversational trees/branches, navigate alternative sibling replies, and execute context-aware generation.
Teaches how to import, parse, and process PDF, DOCX, PPTX, and image files using anchor-mapped and OCR import modes in LollmsDiscussion.
Teaches how to coordinate multiple specialized agents, distribute tasks, establish consensus, and manage role-based multi-agent swarms using LollmsSwarm.
Guides the model on performing basic text generation, counting tokens, detokenizing lists, and generating structurally validated JSON outputs with LollmsClient.
| 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" |
This skill explains how to utilize and interact with the tiered long-term memory system of the LoLLMS framework.
The system prevents context bloating by categorizing information into distinct tiers:
[7c8855bf]) are injected to save token space.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).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.
decay_rate_per_day).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']}")