AIPMAndy/dna-memory

• **Archetype**: utility-with-skill
• **Stars**: 53
• **Last pushed**: 2026-04-01
• **License**: Apache-2.0
• **Discovered**: 2026-04-12
• **Skills found**: 1
• **Source**: gh-search

Summary

A three-tier memory system for AI agents (working/short-term/long-term) with weight-based reinforcement, time-based decay/forgetting, reflection-driven pattern extraction, and memory promotion. Built on SQLite with FTS5 search. Originally designed for the OpenClaw ecosystem but the memory architecture and lifecycle patterns are transferable.

Assessment

The memory lifecycle patterns (remember, recall, reflect, decay, promote, link) are transferable as a process definition. The three-tier architecture with weight decay and promotion thresholds encodes a well-defined state machine. The daemon pattern for background maintenance (auto-reflect, auto-decay) maps to a monitoring/maintenance process.

Repo-specific: Python implementation, OpenClaw-specific installation paths, Chinese-language documentation. The SQLite schema and Python scripts are implementation details -- the *process patterns* are what we want.

Extraction Priority

• Medium
• Rationale: The memory lifecycle (remember -> reinforce -> reflect -> promote -> decay) is a clean process pattern applicable to any agent memory system. However, the babysitter SDK already has some memory/context management patterns, and this would primarily be useful as a specialization process for "agent memory maintenance" rather than a core methodology. The reflection/promotion cycle is the most novel extractable pattern.

---

Processes

(No extractable processes -- memory systems are plugin territory, not process library entries.)

Plugin Ideas

- What install.md would do: Create memory/ directory structure with tier-specific storage, configure decay/reflect intervals in .a5c/config (auto_reflect_interval_minutes: 30, auto_decay_interval_hours: 24), set up session lifecycle hooks, install maintenance scripts for periodic memory hygiene. - Processes it would copy: None (plugin-internal logic) - Configs/hooks it would create: on-session-start hook to load relevant long-term memories; on-session-end hook to filter and persist working memory; periodic maintenance cron for decay/reflect/promote/dedupe cycle; memory type taxonomy config (fact/preference/skill/error/pattern/insight); .a5c/commands/memory-evolve.md slash command for manual maintenance. - Source evidence: SKILL.md three-tier architecture, scripts/evolve.py (reflect, decay, promote, dedupe commands), daemon section with configurable intervals. The weight-based promotion criteria (weight > threshold + stable over time + verified by reflection) and reflection-driven pattern extraction (gather recent high-weight items -> cluster -> synthesize -> store) are the core novel patterns.

• **Agent Memory Plugin** (Context & Memory category): Installs a structured three-tier memory system (working/short-term/long-term) with weight-based decay, reflection-driven pattern extraction, and automatic promotion/demotion.

Library Mapping

Plugin Marketplace Mapping

Implicit Procedural Knowledge

(No extractable procedures -- all memory-related patterns belong in the plugin above.)