thedotmack/claude-mem
Claude-Mem is a plugin for Claude Code that captures, compresses, and preserves coding session context to enable persistent memory and continuity across sessions.
Awesome AI Agents › Continual Learning Agents
aiming-lab/MetaClaw
MetaClaw is a system that lets a personal agent meta-learn and evolve from live conversations instead of offline training alone. It places a proxy in front of the model that intercepts interactions from a personal agent such as OpenClaw, CoPaw, IronClaw, PicoClaw, ZeroClaw, NanoClaw, NemoClaw or any OpenAI-compatible client, injects relevant skills at each turn, and meta-learns from accumulated experience. For Anthropic-native agents it exposes a /v1/messages compatible endpoint so the pipeline works without agent-side changes. Setup is handled by two commands, metaclaw setup for a one-time configuration wizard and metaclaw start to bring up the proxy and wire the chosen agent. Three operating modes are available: skills_only proxies the LLM API and auto-summarizes skills after each session with no GPU or training backend needed; rl adds reinforcement learning with GRPO that trains as soon as a batch is full, with optional teacher distillation; and auto, the default, combines skills and RL with a scheduler that defers weight updates to sleep hours, idle time or calendar meetings so the agent is never interrupted during active use. Support and query sets are separated to keep stale reward signals from polluting model updates. A contexture layer persists cross-session memory for users and projects, retrieving relevant facts, preferences and project history and injecting them into prompts, with adaptive memory policy, background consolidation, incremental ingestion every few turns and an optional memory sidecar service. Serving, reward modeling and training are fully decoupled and asynchronous, so the agent keeps responding while scoring and optimization run in parallel. Training runs on cloud LoRA backends, with Tinker as the reference path and MinT or Weaver available through compatibility packages. The project ships as a native OpenClaw extension and is documented in an accompanying technical report.
https://github.com/aiming-lab/MetaClaw
Also in Continual Learning Agents
vectorize-io/hindsight
Hindsight is an agent memory system designed to create smarter, learning AI agents, addressing limitations of traditional memory techniques like RAG and knowledge graphs.
langchain-ai/langchainjs
LangChain.js is a TypeScript framework for building context-aware, reasoning applications powered by large language models, offering modular components, chains, agents, and tools for production and deployment.
aden-hive/hive
Hive is an outcome-driven agent development framework designed for evolving agents with AI evaluation, human-in-the-loop, and self-improvement capabilities.
kayba-ai/agentic-context-engine
Agentic Context Engine (ACE) is a framework that enables AI agents to autonomously learn from their successes and failures, continuously improving their performance across various tasks without additional training data or fine-tuning.
AgentDock/AgentDock
AgentDock is an open-source framework for building sophisticated AI agents with configurable determinism, enabling reliable and creative AI applications through a node-based architecture and multi-stage workflows.
modelscope/AgentEvolver
Self-evolving agent training framework from ModelScope that combines automatic task generation, experience-guided exploration and attribution-based credit assignment into one reinforcement learning stack.
ghostwright/phantom
Self-hosted persistent AI co-worker built on the Claude Agent SDK that runs on its own machine with long-term memory, self-evolution, an MCP server and Slack, Telegram, email and web channels.