Agent frameworks & infrastructure
Agent memory & context
4,032 agents · most starred first
Agent memory & context, page 24 of 41
Model-driven context-window rollover bundle for the DeepSeek Harness: fresh working context, durable notes, verbatim recent tail, no summarization
给 DeepSeek Harness 的三层记忆:常驻事实 + 会话摘要 + 全文检索。注入不花任何模型调用,写回用你当前选的模型。
Long-term memory for AI agents on Jev. Keep raw records, judge them with a fast System 1 model and answer from under 4k tokens of context.
DSH 上下文智能压缩 + 智能思考:每轮注入用量、模型自主决定压缩并自动续跑;可选让模型按任务难度自主调整思考强度档位
🧠 The definitive curated map of memory for LLM agents — frameworks, research papers, benchmarks, taxonomy & deep dives.
The self-hosted bureau for AI agents: they get briefed on missions from a durable task queue, file what they learn into a markdown+git brain, and show up for work in a 8-bit Gameboy inspired office.
Pre-action evidence revalidation for AI agents. Detect stale reasoning before agents write, send, update, or delete.
Verified-lesson memory for AI agents: records runs, keeps only what your own check passed, and measures with a built-in holdout whether it helps. .NET/MAF, Python, Go, MCP + CLI (aexp).
Local-first context compression and exact recovery for Codex CLI and Claude Code CLI on macOS, Linux and Windows.
Logos is a tool which allows all your ai agents to share the same context and pick up right after one another
DeepSeek Harness 的上下文压缩插件:自动与手动压缩、工具结果裁剪、有界溢出恢复。
Open-source field experiment in persistent AI-agent identity, memory, creation, and interaction. Run the reproducible 10-minute protocol.
Self-hosted AI memory and RAG layer with local LLM support. Full Rust implementation with vector embeddings, knowledge graph, FTS5, hybrid search, and OpenAI-compatible API.
A decentralized cooperative memory & research layer for AI agents — collectively and cooperatively learning and advancing as a community.
岁月涟漪 · Ripples of Aion — A structured memory layer plugin for Cyrene.
A cognitive persistence protocol for multi-agent AI systems — hash-chained, auditable records of how agent state evolved, with provenance and governance rules in the specification.
The Knowledge Lifecycle of Large Language Models: a unified five-stage framework for how LLMs acquire, store, retrieve, update, and forget knowledge, with the Provenance Vector proposal and the reproducible Lifecycle Desynchronization metric.
Four-layer memory substrate for long-horizon agents: write-gate admission, signed approvals, hash-chained ledger, revocable and portable.
Save and use agent configs via sqlite
Text-blind context control plane for AI agents. Select high-utility context under token budgets.
Open-source long-term, time-aware memory engine for AI agents (MCP + hybrid search + bi-temporal recall)
Local-first encrypted memory for AI agents
Research into a new method of context management
🧠 Empower Claude with memory to enhance code debugging and conversation recall, turning ephemeral chats into lasting insights.
Persistent belief-graph memory for AI agents. Retrieves decisive context by importance — not recency, not RAG, not /compact.
A modular, extensible Rust-based server providing short-term, long-term, and latent memory services, a chat endpoint with a BaseAgent + Sentience DSL, and integration with ChromaDB and LLM services.
remem provides agents with persistent, reasoned memory that spans across sessions. Whether you're using Claude Code, Codex, Cursor, Copilot, Antigravity CLI, OpenCode and more.
🧠 Local memory plugins for OpenClaw — checkpoint injection, Qdrant semantic recall, auto-capture. Your memories stay on your machine.
CORTEX for Python -- The World's #1 Memory Architecture for AI • Memory that thinks, learns, and dreams • Synthetic cognition infrastructure for Agentic Ai. pip install cortex-ai
Local memory for the models you already use: hybrid search, encrypted at rest, unlock to work. Only what you send each turn goes out—Claude (MCP), ChatGPT (Actions), same vault.
OpenCxMS Memory System - Persistent memory for AI coding assistants through structured documentation
Autonomous memory and learning system for AI coding agents. LLM-powered knowledge extraction, triple fusion retrieval (FTS5 + TF-IDF + entity graph), quality-gated storage, and self-curating consolidation. Your agent never makes the same mistake twice.
Cross-tool memory for AI agents. Stitch context across Cursor, Claude Code, Codex, Gemini CLI, Windsurf, Copilot, and 9+ AI tools. Preserves decisions, failed experiments & warnings across sessions. Zero dependencies, MCP + instruction files, BM25 relevance search. Never re-explain your work.
Semantic memory API skill for AI agents
Cognitive Ingestion & Belief Formation Engine. Long term AI memory.
Temporal Layered Context Memory (TLCM). The architecture for AI that treats memory as a living timeline instead of a flat filing cabinet.
Lightweight persistent memory for AI agents. SQLite, zero mandatory dependencies, privacy controls (forget/TTL), full audit log. Drop into any agent in 3 lines.
🧠 Synapse - Supercharge your AI coding assistants with memory, context, MCP tools, and intelligent routing. Works with Cline, Roo, Cursor, and any OpenAI-compatible tool. Just change the API endpoint and get superpowers! ✨
Build an AI Agent with long term memory using LangGraph
The privacy-first memory engine for AI agents. Setup wizard, per-user vaults, smart inbox, natural-language forget, MCP server for Claude/Cursor/Windsurf. Zero API calls. Zero dependencies. Pure Python.
AutoGen Multi-agent Conversation and Recall using Memori
The serialization engine for LLM agentic workflows. A drop-in Python/Rust replacement for JSON that mathematically reduces context window bloat by 44% and deserializes 3x faster. Built for multi-agent swarms, RAG memory, and AI orchestration.
Offline NPC memory middleware — 16-byte factor vectors, SQLite only, no cloud. NPCs remember, forget, and dream.
🧠 Universal long-term memory for AI agents. GraphRAG-powered knowledge base with vector search + graph traversal. Privacy-first, local-only, MCP-compatible. Connect Claude, Copilot, or any AI assistant.
A toy local memory system for multi-agent workflows
ION: Context engineering & agent governance
A cognitive runtime architecture for LLM continuity, memory, identity, and safe pseudo-conscious agents.
An airline customer service AI agent demo built on AWS Bedrock AgentCore with Strands Framework, featuring Gateway policy enforcement and Memory (STM/LTM). 基于 AWS Bedrock AgentCore 构建的航空客服智能代理演示系统。
Recollect is a memory layer for AI agents that provides robust context compaction for long-running chats.
A powerful, pluggable vector memory and Model Context Protocol (MCP) server for local semantic search and long-term memory.
Kodama for Claude Code — Reliable session memory & smart restart for the Claude CLI
Layered memory system for LLM agents using semantic, episodic, and procedural memory with Qdrant-backed long-term storage.
An agent that Knows You, Remembers You, and Understands You
CCMem - Claude Code Memory: Standardized MCP server providing persistent project memory and context-aware development assistance for Claude Code
Never lose context again. Structured memory system for Clawdbot/Moltbot. Built by Nox ⚡
AI 助手的模組化能力框架:記憶、防禦、診斷、品質穩定 | Modular capability framework for AI assistants | Claude Code / Cursor / Any LLM
Seamless, infinite context for your LLMs. Handles long conversations and long messages.
Give Claude Code a memory that persists across sessions and survives context compaction. Zero dependencies, pure Python, 30-second install.
Connection-based memory lifecycle for Claude Code — entries that connect to other knowledge survive, isolated ones fade. Built entirely with hooks.
Persistent identity and memory across AI tools - mcp-native, local-first, framework-agnostic, production-ready.
Hands-on workshop to learn how to implement context engineering with Java, LangChain4J, and Redis
Memlane is where you can store you memories that you think you will forget, but it never
CLO (Cross-LLM Orchestrator) is a specialized browser extension designed to eliminate context loss when switching between different LLM platforms. It acts as a **"Reasoning State Virtual Machine"**, capturing the cognitive state of a conversation and enabling it to be "rehydrated" into any other model seamlessly.
A Go LLM client with a memory. Every provider, cost-aware, single binary.
AI browser memory for anyone who works in a browser. Local-first, private, BYOK.
Enable AI agents to share memory across tools and avoid repeating work when switching between tasks or platforms.
Mímir's Well — AI memory database with Ebbinghaus decay, self-healing, backup rotation, and full-text search
A cognitive runtime that gives LLM agents persistent state, identity, and learning across turns. Memory, beliefs, drives, self-evolution, skills, and affective state — engineered scaffolding outside the model.
Reusable design patterns for long-running LLM agents with stable identity and controlled change.
Collective intelligence layer for AI agents — search what others already solved, publish what you learn, never start from zero.
Tamper-evident memory & decision lineage for AI agents. Cryptographic proof of what your agent knew. Powered by a hybrid Python/Rust (Direct-Silicon) core.
Persistent graph memory architecture for multi-agent AI systems using Graphiti, FalkorDB, and MCP.
Eidentic is the open-source TypeScript SDK for AI agents with self-improving memory and production fundamentals built in.
Memory OS AI — Adaptive memory system for AI agents (forked from RoaminS/Memory-os-ai)
Simple node based memory system for AI agents. Built in Swift. CLI, MCP (stdio and HTTP), and observation WebSocket server.
Memory migration tool for the Claw AI agent ecosystem — WorkBuddy, OpenClaw, QClaw, CoPaw and more
Enable AI agents to retain memory across sessions using persistent storage designed for continuous context retention.
BrainRouter is a cognitive memory and multi-agent orchestration platform for AI coding agents. It provides a first-party terminal CLI, MCP-based memory tools, layered recall, context compaction, graph memory, and dashboard inspection workflows to help agents preserve long-running task continuity while reducing prompt size and inference cost.
Token-efficient context engine for AI coding agents. v2.23.0: token-killer pack - deep-forget (unlearning propagation), dangling-reference guard, MCP schemas-on-demand. 2694 tests, 43 MCP tools, Elastic-2.0.
Cairn (AI agent recovery library) v1.0.0: a framework-agnostic Python library and benchmark that lets long-horizon AI agents recover from crashes by re-grounding, not replay.
MCP memory server for Claude Code, Cursor and AI agents - Rust local-first AI memory with hybrid search, SQLite, knowledge graph and Python/TypeScript SDKs
Reduce token usage for JSON, CSV, and TSV files with deterministic compression that keeps data intact for AI agents.
The open-source memory for AI agents — it gets smarter the more it remembers, not slower or costlier, and never confidently remembers wrong. Plain files, no database, bring your own model. Works with Claude Code.
Turn any conversation into a portable knowledge graph. TRON + Sparse6 + psychology — paste into any AI.
Long-term memory your AI coding agents actually share. One brain, many agents — standalone embedded MCP server: SQLite + FTS5 + vector hybrid search, local ONNX embeddings, self-expiring diary. No server, container, or daemon.
Reference implementation of the Bitemporal Provenance Memory Architecture (BPMA): an append-only, bitemporal, provenance-first memory ledger for AI agents.
A knowledge management platform which acts as a persistent, structured, and token-efficient memory layer for large language models
Reduce token use and enable persistent memory for Claude Code, boosting efficiency in solo or team agent setups.
ARCHIVED — historical Hermes/Headroom integration experiment; use the upstream retrieval-only plugin as baseline.
Lazy-loaded context management for LLM external calls. Archive tool/skill/MCP results after each turn, replace with compact stubs, and give the model a query tool to retrieve past results on demand. Context stays lean — nothing is lost.
Context Memory eXchange — a retrieval-first, engine-enforced-grounding context engine for Hermes Agent and a drop-in LCM replacement. Keeps 100% verbatim history (SQLite FTS5+trigram+embeddings; Postgres optional), retrieves exact slices every turn, and forces + verifies grounding so the model answers from real history or honestly refuses.
Context Layering & Engineering for Agentic Resources — Persistent memory, intelligent context, and structured project management for Claude Code
Drop-in compression for LLM agent tool outputs: cut ~99% of tool-result tokens while keeping (often improving) answer accuracy. Provider-agnostic, with a faithfulness benchmark and an OpenAI-compatible proxy.
A recipe for any AI agent to build a self-improving theory-of-mind model of its user from interaction logs (feedback precognition). What you mind is what you get.
An open framework for a persistent, self-regulating AI operations system on a coding-agent CLI: tiered memory + SSOT, security-first gates, a knowledge pipeline, multi-agent protocols, session survival, and a zero-dependency cockpit.
A production-style AI assistant with real-time streaming, semantic memory (STM + LTM), and contextual conversations powered by Gemini, Pinecone, and Socket.IO.
Chat history search for your llm
Lazy-loaded context management for LLM external calls. Archive tool/skill/MCP results after each turn, replace with compact stubs, and give the model a query tool to retrieve past results on demand. Context stays lean — nothing is lost.
Autonomous research org shipping measured, reproducible results on AI agent memory & reasoning — a public replication ledger (the Crucible), the RAMR benchmark, and mnemo, a zero-dependency memory core. Every claim ships with runnable code that could prove it wrong.
Agent-first protocols for LLM memory, active workspace, skill lifecycle, councils, and stewardship—with explicit maturity, evidence, and conformance.