Memory AI apps

Memory graph and retrieval layer that gives AI agents long-term context.

Papr is a memory layer for AI agents that turns unstructured data—documents, conversations, and logs—into a searchable knowledge graph with long-term recall. It offers graph-aware retrieval, automatic chat-memory compression, and TypeScript and Python SDKs, exposed through a REST API and a managed cloud dashboard. Its papr-2 model reports sub-150ms cached retrieval, and the core can be self-hosted.

SQL-native memory engine for LLMs and AI agents.

An LLM-agnostic memory layer that turns agent conversations and execution into persistent, queryable state stored in standard SQL databases (Postgres, MySQL, SQLite). It adds long-term memory to any LLM app with roughly one line of code, using a dual-mode system for working and long-term recall. MemoriLabs also offers a hosted enterprise memory fabric on top of the open core.

A 'visual memory' layer for AI — search and reason over huge video libraries.

Video understanding platform built around a large visual memory model. It ingests long-form and large-scale video, then supports natural-language search, transcription, clip retrieval, and content analysis with unlimited video context. Applied to security and surveillance review, sports analytics, media production, and robotics, with a free playground and on-device processing options.

Open-source agent memory that retains, recalls, and reflects.

Hindsight is an open-source memory system for AI agents that works like human memory — organizing world facts, agent experiences, entity summaries, and evolving beliefs into four networks. Its retain/recall/reflect operations let agents reflect on past runs and update beliefs instead of repeating mistakes. It was built by Vectorize with Virginia Tech and The Washington Post.

Portable single-file memory layer for AI agents — no database.

Memvid stores an agent's knowledge as compressed 'Smart Frames' in a single portable .mv2 file that bundles data, embeddings, and the search index — replacing vector databases and complex RAG pipelines. It offers sub-5ms retrieval, append-only versioned storage, and runs fully offline. The format is model-agnostic and multimodal.

Self-evolving long-term memory OS for AI agents.

EverMind builds persistent memory infrastructure for AI agents. Its EverOS (a.k.a. EverMemOS) is a brain-inspired memory operating system — sensory encoding, hippocampal-style indexing, and long-term consolidation — that gives agents context that survives across sessions and platforms, with multimodal retrieval and agent-trajectory recording. It's a portable layer meant to plug into Claude Code, Codex, and other agents; self-host the Apache-2.0 core or use the managed EverOS Cloud.

Open-source memory layer for AI agents.

MemMachine is an open-source memory layer that lets AI agents learn, store and recall information across sessions, turning stateless chatbots into personalized, context-aware assistants. It combines episodic memory (conversational context) with profile memory (long-term user facts) and is model-agnostic, working across providers. It is self-hosted and built primarily in Python.

Memory and context layer for AI agents.

Hyperspell is a context and memory layer that connects an agent to its user's workspace data — Slack, Gmail, Notion, Linear, Salesforce, and more — through 50+ pre-built connectors. It continuously indexes that data into a context graph and exposes it to agents as a queryable filesystem, so any framework can give its agents governed, personalized memory. It targets production use with SOC 2 and GDPR compliance.

Memory framework for long-running, always-on AI agents.

memU is a memory framework for proactive, always-on AI agents from NevaMind AI. It ingests multimodal inputs (conversations, documents, images), extracts them into structured memory organized as a hierarchical file system, and retrieves the right context via vector-indexed semantic recall. The approach is designed to keep agents online continuously at a fraction of the usual token cost. It is open source under Apache 2.0, with self-hosting plus a hosted cloud API.

An always-on AI second brain that captures and remembers everything you hear.

TwinMind runs in the background on your phone, continuously listening to meetings, lectures, and conversations to build a private, searchable memory you can query in plain language. It transcribes on-device, generates notes, to-dos, and daily digests, and answers questions across everything it has heard. The proprietary Ear-3 speech model handles 140+ languages, and audio is processed locally for privacy, with cloud AI chat in the paid tiers.

A private, AI-organized place to remember everything.

mymind is a personal memory tool: save notes, links, articles, images, and quotes, and it auto-tags and sorts them for you — no folders, no manual filing. AI recognizes what each item is (article, product, recipe, book) and surfaces it again through fast visual, associative search. It's built privacy-first, with no ads, no tracking, and nothing sold or used to train models.

Persistent memory layer for AI coding agents.

ByteRover gives coding agents long-term, structured memory that persists across sessions and tools. It organizes project knowledge into hierarchical context trees, stores them locally by default, and serves relevant context back to agents through an MCP server and CLI. It plugs into Cursor, Claude Code, Windsurf, and other assistants, and an optional cloud tier lets teams sync and share context.

Persistent, local knowledge graph for AI tools — stored as plain Markdown.

An MCP server that gives AI assistants a durable, local knowledge base built from your conversations, notes, and documents. Everything is written to standard Markdown files you own, indexed by a local SQLite store, and readable/editable in Obsidian. Connects to Claude, Cursor, VS Code, and other MCP-compatible clients.

Long-term memory SDK that lets agents learn and adapt over time.

A memory toolkit for AI agents that extracts and stores facts from conversations and refines agent behavior through prompt updates. Exposes semantic, episodic, and procedural memory via a storage-agnostic core API plus tools agents can call during or after a conversation.

Local-first 24/7 screen and audio memory your AI agents can search and act on.

Screenpipe continuously captures everything you see, say, and hear on your computer — extracting text via accessibility APIs and OCR and transcribing audio locally — to build a private, searchable memory of your digital life. That context feeds AI agents and 48+ app integrations, with on-device PII removal and all data stored locally by default. Cloud sync and managed deployment arrive in the paid tiers.

One API for AI agent memory: ingest, extract, store, retrieve.

A cloud-native platform that gives AI agents semantic memory and operational context through a single API. It ingests documents, audio, video, and web pages, extracts entities and relationships, and handles storage and retrieval so you don't assemble the RAG pipeline yourself. Integrates with frontier models from OpenAI, Anthropic, and Google for extraction.

On-device AI memory and copilot that recalls your work context.

Pieces is a developer-focused AI tool that automatically captures your work context — code snippets, docs, chats, and links — across your apps and surfaces it through a copilot. Its Long-Term Memory engine records a rolling window of activity at the OS level, enabling time-based questions about what you were doing. It runs on-device and air-gapped from the cloud by default, with optional cloud LLMs, and plugs into desktop, VS Code, JetBrains, and the browser.

Open-source memory for AI agents.

An open-source semantic memory layer for AI agents. Cognee ingests documents, relational data, and system context, then runs an Extract-Cognify-Load pipeline that uses an LLM to build a knowledge graph with embeddings and relationships. Agents query it for durable, cross-session context that captures how concepts connect. Self-host the Python SDK for free, or use the managed cloud tiers.

Continual learning memory for stateful agents. Better context, fewer tokens.

A memory and user-personalization layer for AI agents that keeps reasoning about each user across sessions, so apps get richer context without stuffing whole histories into the prompt. It models peers and sessions, runs background inference to derive durable facts, and answers natural-language questions about a user at query time. Available as a managed API or self-hosted FastAPI server, with Python and TypeScript SDKs.

User profile-based long-term memory for LLM applications.

Memobase gives AI apps persistent, per-user memory by distilling conversations into a structured user profile and a chronological event timeline, rather than storing raw transcript embeddings. It exposes Python, Node, and Go SDKs plus a REST API and an MCP server, and is open source (Apache-2.0) with a managed cloud option. Built on FastAPI, Postgres, and Redis for low-latency retrieval.

Temporal knowledge-graph memory for AI agents.

Memory layer that gives agents long-term context by building a temporal knowledge graph from chat history and business data, tracking how facts evolve over time. It's powered by Graphiti, Zep's Apache-2.0 open-source temporal graph engine, with Zep Cloud offering a managed, credit-based service on top. Used to keep agent context relevant as conversations and data grow.

Memory API that gives any AI agent long-term recall.

Supermemory is a memory and context engine for AI apps. It ingests documents, chat histories, and connector data (Drive, Gmail, Notion), turns them into a searchable store, and serves relevant context back to agents over a single API. It works with any model and ships an MCP server alongside official SDKs.

Self-improving personal AI agent that learns skills and keeps persistent memory.

A self-hostable AI agent from Nous Research that builds skills from experience, maintains long-term memory, and is reachable through a terminal TUI, a web dashboard, or messaging gateways (Telegram, Discord, Slack). It runs anywhere from a small VPS to local Docker, and works with any model via Nous Portal, OpenRouter, OpenAI, or self-hosted endpoints. Free and open source under MIT; you bring your own model access.

Long-term memory layer for AI agents. Self-hostable.

Persistent memory store + retrieval pipeline for agent applications. Handles per-user/per-session/per-agent scope cleanly; pairs with OpenAI, Anthropic, and local models.

Stateful agents with structured memory. Successor to MemGPT.

Open-source framework for building stateful agents — memory blocks, context-window management, tool-use primitives baked in. Useful as a reference architecture for long-running agents.