Over
files, workspaces, and memories created
What is Exabase Memory?
Exabase Memory is a self-organizing memory engine for AI agents. Send in conversations or raw text and the M-1 engine extracts facts, preferences, and events automatically.
But it doesn't just store them. It builds a living ontology: tracking relationships between concepts, resolving contradictions as facts change, and consolidating knowledge over time.
Not a vector database. Your agent's memory gets more accurate the more it's used, not noisier.
Problem
Agents can't remember anything.
Your agent forgets everything between sessions. Users repeat themselves. Your copilot suggests the same rejected patterns. Support bots ask questions they've already answered.
You try storing conversation logs and running vector search, but it falls apart with conflicting facts, stale context, and noise.
Memory is harder than it looks.
Solution
Memory that actually works
Exabase Memory isn't just storage and retrieval. It tracks relationships, handles contradictions, and evolves over time.
Your knowledge base gets better over time, not worse.
Customer support example:
A user says they are on the free plan. They upgrade to enterprise two weeks later. The support agent still retrieves and reasons from their original plan tier, offering limitations that no longer apply and missing features they now have access to.
Exabase Memory knows the relationship changed.
Datapoint
Exabase M-1 memory system is benchmarked as #1 in the world on the leading benchmark.
Automatic memory extraction
Send your agent's conversations to Exabase. It extracts facts, preferences, and events automatically.
No need to decide what's worth remembering.
Or set memories directly
Write specific memories through the API when you need precise control.
Directly set any text content as a memory.
Retrieve memories
Code:
Response:
Hybrid semantic and keyword search with recency bias.
Exabase can expand your query into multiple retrieval passes and rerank candidates by relevance.
Control precision with expandQueries (up to 5), rerankCandidates (up to 500), and rerankThreshold (0-1).
One call returns everything relevant for a task, ranked and ready for your prompt.
Memory that improves accuracy
Memory isn't just about personalization.
It's about reliability.
Agents hallucinate when they either lack or have too much context.
They make up facts, forget constraints, ignore preferences the user already stated. Memory grounds your agent in what actually happened.
Exabase Memory updates continuously as new information comes in. Contradictions get resolved. Stale context gets refreshed. Your agent stays accurate, not just personalized.
28% fewer hallucinations when agents have access to reliable memory and context.*
*Based on research into context-augmented AI systems.
Use cases
Why Exabase
Part of a full platform
Related posts:
FAQs
What is Exabase?
Exabase is infrastructure for AI agents. It gives your agents memory, versioned file storage, AI deep search, and context automation through a set of APIs. Store what your agent learns, search inside any content type, and keep knowledge bases current automatically. Built for production use.
Who uses Exabase?
Developers and teams building AI agents, copilots, and RAG applications. If your agent needs to remember things between sessions, store and retrieve files, search across documents and media, or stay up to date without manual maintenance, Exabase handles that infrastructure so you can focus on your product.
What is Exabase Memory?
Exabase Memory is a self-improving memory system for AI agents. It stores facts, preferences, and events from conversations, organizes them automatically, and makes them searchable via semantic search. It's built on the M-1 engine, ranked #1 on the LongMemEval benchmark.
How does Exabase Memory differ from a vector database?
Vector databases store embeddings and return results based on cosine similarity. Exabase Memory builds a dynamic ontology of concepts and relationships. It tracks how facts relate to each other, resolves contradictions when information changes, and self-organizes over time. It's a memory system, not a search index.
What is the M-1 engine?
M-1 is Exabase's proprietary memory engine. It handles extraction, organization, contradiction resolution, and retrieval. It's benchmarked as #1 on LongMemEval, the leading evaluation for AI memory systems.
How do I add memories?
Two ways. Send raw text with source: 'text' and Exabase infers structured memories automatically (inferred mode). Or set infer: false to store exact text as a verbatim memory. Inferred memories are async. Verbatim memories return immediately.
How does memory search work?
POST to /v2/memories/search with a natural language query. Exabase uses hybrid semantic and keyword search with recency bias. Results are ranked by relevance, each with a score.
Can I scope memories per user?
Yes. Use Bases to isolate memories by user, project, or context. Pass the X-Exabase-Base-Id header with any memory API call. Each Base is a fully sandboxed container.
Does Exabase handle contradictions?
Yes. When new information contradicts existing memories, Exabase resolves the conflict automatically. If a user moves from Toronto to Berlin, the memory system updates accordingly. Old facts don't pollute new context.
Can I make memories permanent?
Yes. Set the immutable flag when creating a memory. Immutable memories won't be modified, consolidated, or deleted by the system's self-organizing process.
How is memory organized internally?
Memories self-organize into a tree-like structure of concepts and memories. Concepts are high-level categories (like "Work" or "Preferences"). Memories are individual facts associated with the most relevant concept. This structure evolves automatically.
What does "self-improving" mean?
Every time new information comes in, Exabase re-evaluates existing memories. It consolidates duplicates, resolves contradictions, recategorizes when topics shift, and prunes stale content. The memory graph gets more accurate and useful over time without manual maintenance.
How is Exabase Memory different from Mem0?
Mem0 provides memory storage and retrieval. Exabase Memory goes further with a proprietary self-organizing engine (M-1), automatic contradiction resolution, and a full infrastructure platform around the memory layer (files, search, extraction, workers). M-1 is #1 on LongMemEval.
How is Exabase Memory different from Zep?
Zep focuses on conversation memory with entity extraction. Exabase Memory builds a dynamic ontology, not just entity-attribute pairs. It resolves contradictions, self-organizes, and integrates with Exabase's full platform for file storage, search, and autonomous maintenance.
What's the response time for memory retrieval?
200ms for simple retrieval. Fast enough to use in real-time agent loops without adding noticeable latency.
Can Exabase Workers manage memories?
Yes. Workers have access to memory tools. You can set up a Worker to create, update, and resolve memories on a schedule. Useful for keeping memory current based on new documents or external data.
Is memory search scoped to a Base?
Yes. If you pass a Base ID, search only returns memories within that Base. If you don't pass a Base ID, search runs against your primary base's system-managed memory section.
What happens if I don't specify a Base ID?
Memories are stored in a system-managed section of your primary base. This works for personal or single-tenant use. For multi-tenant applications, create a Base per user or project and scope all memory operations to it.
