Memexai
we gave AI agents a "dreaming" phase to consolidate memory
Last verified:
What is Memexai?
MemexAI is a persistent memory infrastructure for AI agents that gives agents durable, inspectable memory without requiring a vector database. Memory is stored as scoped Markdown files in Postgres, with revision history, access logs, and a self-hosted admin UI for human inspection. The system introduces a 'dreaming' phase that mimics human sleep-based memory consolidation, where the AI reorganizes, consolidates, and compresses memory traces during idle periods to improve recall accuracy.
Key features include two default tools: memory_memorize (extracts and writes durable facts) and memory_search (recalls memory via BM25 full-text search with optional LLM-assisted resolution). The platform supports multiple memory types—episodic, semantic, and procedural—and offers both direct Postgres integration (@memexai/core) and HTTP service (@memexai/sdk + Docker). Adapters for Vercel AI SDK, Anthropic SDK, LangChain, and OpenAI are included. Memory visualization tools help developers understand how agents organize and recall information.
MemexAI is designed for developers building AI products that talk to users across sessions, teams who need to correct wrong facts agents are carrying, and anyone whose users complain that 'the AI forgot what I told it.' It is best for building persistent AI agents with long-term memory, multi-session conversational applications, research and development of AI memory systems, and complex task automation requiring context retention.
Memexai pricing
Pricing model: Freemium
Freemium pricing model. Free tier includes 10,000 memory operations per month with basic consolidation features. Pro tier at $29/month unlocks unlimited operations, advanced dreaming algorithms, and priority support. Enterprise plans offer on-premise deployment, custom memory architectures, and dedicated engineering support.
Memexai pros
- No vector database required - uses Postgres instead
- Memory stored as inspectable Markdown files
- Revision history for every memory change
- Access logs track all memory reads
- Self-hosted admin UI for human inspection
- Dreaming phase consolidates memory like humans
- BM25 full-text search built-in
- Optional LLM-assisted resolution for complex queries
- Supports episodic, semantic, and procedural memory
- Pre-built adapters for major AI SDKs
- Direct Postgres or HTTP service integration options
- Scoped virtual paths ensure agent isolation
- Memory visualization tools for developers
- Pre-built memory schemas for common use cases
- Generous free tier to get started
Memexai cons
- Early stage product with limited maturity
- No code demo available to evaluate
- Dreaming phase requires idle time to process
- Limited to Postgres database
- Raw toolset requires explicit file control knowledge
- Self-hosting requires Docker setup
- LLM-assisted resolution requires model configuration
- Free tier has 10,000 memory operations monthly limit
Frequently asked questions about Memexai
What is MemexAI and what problem does it solve?
MemexAI gives AI agents durable, inspectable memory without a vector database. It solves the problem where most agent memory lives in chat history or hidden app tables, causing agents to forget user preferences across sessions. When a user says 'I prefer quiet neighborhoods' in session 1, the agent shouldn't forget in session 2, and developers should be able to audit what the agent remembered and when.
How does the dreaming phase work?
The dreaming phase is a biomimetic memory processing that occurs during idle periods. During this phase, the system reorganizes, consolidates, and compresses memory traces, strengthening important connections while pruning irrelevant information. This dramatically improves recall accuracy and enables more meaningful information synthesis across disparate data points.
Why does MemexAI use Postgres instead of a vector database?
Most agent memory is structured recall, not nearest-neighbor search: what does this user prefer, what has the agent already learned, what changed and who changed it? Postgres covers storage, full-text search, migrations, access control, and audit in one place. Embeddings are optional.
What tools do agents get by default?
Agents get two tools by default: memory_memorize which extracts and writes durable facts, and memory_search which recalls relevant memory via BM25 full-text search or LLM-assisted resolution if a model is configured. A raw toolset is also available for explicit file control.
What integration paths are available?
Two integration paths are available: Direct Postgres using @memexai/core which imports and connects with no HTTP service needed, and HTTP service using @memexai/sdk plus Docker which provides API and admin UI in one container. Adapters for Vercel AI SDK, Anthropic SDK, LangChain, and OpenAI are included.
How can humans inspect and fix memory?
Memory is stored as scoped Markdown files that humans can open and edit. If a memory is wrong, you can open the file, edit the line, and the next response changes. The admin UI shows what was remembered, when it changed, who touched it, and which reads happened later. Every change has revision history and access logs.
What memory types does MemexAI support?
MemexAI supports multiple memory types including episodic memory for conversation-level recall, semantic memory for long-term knowledge and user preferences, and procedural memory for learned behaviors and patterns. This allows developers to build agents with human-like memory capabilities.
How do I get started with MemexAI?
You can clone the repository and run docker compose up -d to get API and admin UI at localhost:8080. The npm packages are @memexai/core for direct Postgres and @memexai/sdk for HTTP. You can also follow the agent onboarding at memexai.space/setup.md which lets your coding agent inspect your app and wire memory tools.
What use cases is MemexAI best for?
MemexAI is best for building persistent AI agents with long-term memory, multi-session conversational applications where users complain the AI forgot them, research and development of AI memory systems, customer service bots that remember past interactions, research assistants that synthesize information across documents, personal AI companions with genuine continuity, and autonomous agents performing multi-day tasks.
How does MemexAI compare to Mem0, Zep, or vector databases?
The important question is whether you need search over old text or an inspectable user memory record your team can operate. MemexAI focuses on durable memory you can inspect and fix, not just RAG over transcripts. Memory is legible before it is clever - using prompt-block injection, Postgres, files, revisions, access logs, and a small set of tools to make memory dependable.