Collection · 34 entries
Memory and knowledge graphs for AI agents
Memory layers, knowledge graphs, and persistent context stores for agents - the substrate underneath useful long-running systems.
'Memory' in agent systems is overloaded. The picks here split between session/portable memory (agentic-stack - a portable .agent folder) and knowledge-graph-backed retrieval (M-flow, codebase-memory-mcp) where the substrate is a graph not just a vector store. Both directions matter for different reasons: the first for working across tools, the second for the relevance-over-similarity problem in RAG.
mcptube - Karpathy-style LLM wiki for YouTube
MCP server that turns YouTube videos into a persistent, merging wiki rather than ephemeral vector chunks. Scene-change frame extraction + vision analysis captures slides, code, and diagrams that transcripts miss. 25+ MCP tools, FTS5+LLM hybrid retrieval, version history with source attribution per claim.
agent-session-resume - cross-agent session resume skill
Skill that lets Claude Code, Codex, Antigravity, and OpenCode pick up where any of them left off by reading and writing a shared session-state file.
Egregore - shared memory for multiplayer Claude Code
Local MIT substrate that gives a team of Claude Code sessions a shared memory and coordination layer, spun up via npx create-egregore.
wuphf - Karpathy-style LLM wiki for agents
Local wiki layer where agents read and write Markdown under git, indexed with bleve (BM25) and SQLite. Designed so context compounds across sessions instead of being re-pasted.
pocket-ingest - voice memos to Claude Code via iCloud
iOS Shortcut records and transcribes a voice memo, drops a .txt into iCloud, and Claude Code picks it up automatically on the next session. No servers or API keys.
engram - context spine for AI coding agents
Aggregates six providers (graph, git, MemPalace, Context7, Obsidian) into rich context packets that fire on every agent Read. Local SQLite, no native deps, no cloud.
witchcraft - multi-vector semantic search in Rust
From-scratch Rust port of Stanford's XTR-Warp running on a single SQLite DB with 21ms p95 latency. Ships Pickbrain, a Claude Code/Codex skill that searches across all your past agent sessions.
stash - persistent memory layer for AI agents
Self-hosted single-binary Go service that stores episodes, facts, and working context for AI agents in Postgres. Ships an MCP server and runs without cloud dependencies.
mex - persistent project memory for AI agents
Structured scaffold and drift-detection CLI that gives Claude Code, Cursor, and other coding agents a project-level memory file that stays in sync with the codebase.
Wax - on-device RAG memory layer in Swift
Single-file Swift memory layer with sub-millisecond RAG on Apple Silicon via CoreML and Metal. No server, no API, ships an MCP server out of the box.
ctx - cross-agent /resume for Claude Code and Codex
Local SQLite-backed skill that stores work as a persistent stream of sessions, notes, decisions, todos, and resume packs. Switch between Claude and Codex without losing the thread.
entroly - self-evolving repo context compressor
Rust daemon that compresses 2M-token repos down ~95% into a Principal Engineer-style context for Cursor, Claude Code, Opus, Codex, and custom providers.
claude-historian-mcp - search past Claude Code chats
MCP server that indexes Claude Code conversation history and exposes it as a searchable tool. Lets agents recall what was decided in past sessions instead of re-deriving it.
graymatter - persistent memory for coding agents
Go CLI that gives any AI agent persistent local memory in three lines, claiming up to 90% reduction in token use for long-running tasks. Offline-first.
mulch - growing expertise files for coding agents
Structured expertise files that live in git and accumulate over time, giving any coding agent a persistent set of project-specific lessons learned.
opencode-magic-context - context cache for OpenCode
OpenCode plugin that adds cache-aware infinite context, cross-session memory, and background history compression. Aims to keep long sessions from blowing past the model context window.
cass-memory-system - shared memory across agents
Procedural memory layer that turns scattered Claude Code, Codex, and Cursor session transcripts into persistent, cross-agent memory. Every agent learns from every other agent's prior work.
auto-memory - progressive session recall CLI
CLI that gives AI coding agents persistent recall across sessions through progressive memory snapshots. Aimed at workflows where context is lost between agent runs.
claude-code-memory-setup - Obsidian-backed agent memory
Persistent memory pipeline for Claude Code using Obsidian and Graphify: codebase knowledge graphs plus a chat import pipeline. Author claims up to 71.5x fewer tokens per session.
trustgraph - context platform for agents
Graph-native infrastructure for storing, enriching, and retrieving structured agent context. Provides semantic retrieval and portable context cores you can move between agent runtimes.
davia - editable docs for coding agents
Documentation system designed to be read and rewritten by coding agents instead of humans. Stores knowledge in a format that survives long agent sessions.
drift - codebase pattern memory MCP server
MCP server and offline CLI that detects codebase patterns and conventions and remembers architectural decisions across sessions, so agents stop drifting from house style.
pro-workflow - self-correcting Claude Code memory
Claude Code workflow where corrections persist as compounding memory across 50+ sessions, with parallel worktrees, agent teams, and 17 prebuilt skills.
MemoMind - local GPU-accelerated memory for Claude Code
Local-first memory system for Claude Code with GPU acceleration and zero cloud dependency. Provides persistent agent memory via MCP.
OpenContext - personal context store for agents
Local context store with desktop GUI for capturing, searching, and reusing project knowledge across Codex, Claude Code, and OpenCode CLIs. Bring-your-own coding agent.
vestige - cognitive memory MCP server
Single-binary Rust MCP server that gives agents long-term memory via FSRS-6 spaced repetition, 29 cognitive modules, and a 3D dashboard. Works with Claude, Cursor, JetBrains.
atlas-mcp-server - Neo4j task system for agents
MCP server backed by Neo4j that gives LLM agents project, task, and knowledge tiers for managing complex multi-step workflows, with deep-research mode.
iwe - Markdown memory graph for AI agents
Rust CLI that turns a Markdown notes folder into a queryable knowledge graph for both you and your agent, with Helix integration and GTD primitives.
memctx - autonomous session memory for Claude Code
Claude Code plugin that captures session summaries, builds searchable history, and auto-injects context into new sessions. Stops you from re-explaining the codebase every morning.
sp-context - persistent agent memory with BM25
Gives every agent persistent memory of team knowledge using Git plus BM25 retrieval - no vector DB, ~114 tokens per session. TypeScript/Bun, works with Claude Code.
claude-memory-compiler - evolving codebase memory
Hooks capture Claude Code sessions, the Agent SDK extracts decisions and lessons, and an LLM compiler organizes them into cross-referenced knowledge articles. Memory that grows with the repo.
M-flow - relevance-first agent memory
Memory layer for agents that surfaces relevant context instead of just similar embeddings. Episodic plus long-term memory backed by a knowledge graph.
agentic-stack - portable .agent folder
One memory + skills + protocols folder that plugs into Claude Code, Cursor, Windsurf, OpenCode, Hermes, or a DIY Python harness - and survives the switch.
awesome-harness-engineering - agent harness toolkit
Curated awesome list for AI agent harness engineering: tools, patterns, evals, memory, MCP, permissions, observability, and orchestration.
Related collections
MCP servers and Model Context Protocol tools
Production MCP servers, gateways, frameworks, and clients - everything in this directory that speaks the Model Context Protocol.
Claude Code tools, plugins, and integrations
The best tools, MCP servers, and harnesses for getting more out of Claude Code - orchestration, observability, telemetry, and remote control.