from dna-memory46
Three-layer agent memory system (working / short-term / long-term) with active forgetting, automatic pattern induction and a reflect loop to let agents learn an
DNA Memory provides a three-layer memory architecture for agents: working memory for current sessions, short-term memory with decay, and long-term memory for consolidated patterns. It adds automated recall, decay, reflection (pattern induction), and linking so agents can remember useful facts, forget noise, and generalise repeated signals into reusable knowledge.
Use when an agent must retain user preferences, learn from interactions over days, summarise repeated events into patterns, or automatically surface past decisions during conversations. Triggers include user mentions of "memory", "remember", "reflect", or any request to persist or recall context.
Best used with agents that support persistent local files and scheduled tasks (Python-based agents, Claude/Copilot-style integrations that can call scripts or read workspace files).
DNA Memory implements a three-layer agent memory system (working/short-term/long-term) with active forgetting, reflection loops, and knowledge graph associations — a solid concept. However, 13 of 14 scripts fail to run due to fragile import paths that assume a specific project directory layout (sys.path.insert hacks pointing to parent dirs, then importing from scripts.*). Only the self-contained evolve.py works standalone. The SKILL.md is comprehensive but entirely in Chinese, limiting global accessibility. The codebase has 14 scripts where many are auxiliary (visualize, trigger, knowme_link) that would be better organized in references/ per the skill spec.
The skill is not malicious — it's a genuine attempt at a multi-layer agent memory system. The main issues are: (1) fragile import structure making 93% of scripts non-functional outside the original repo, (2) Chinese-only documentation limiting adoption, (3) 14 scripts when the skill spec recommends lean SKILL.md with details in references/, and (4) the memory_db.py shim file existing solely to paper over broken imports between scripts. The core concept (three-layer memory with forgetting + reflection + knowledge graph) is valuable and well-documented in the SKILL.md, but the implementation needs structural cleanup to be reliably usable.
