Why Exabase is becoming the default memory layer for AI agents

The agent memory category has grown quickly. Two years ago, most teams were rolling their own on top of a vector database. Today there are dedicated platforms, open-source frameworks, and hosted APIs competing for the same developer. The options are better than they have ever been.

The question is no longer whether to use a memory layer. It is which one, and on what basis to evaluate them.

The benchmark picture

M-1 holds state-of-the-art results on both LongMemEval and BEAM, the two major benchmarks for conversational memory retrieval. No other system holds the top score on both.

On LongMemEval, M-1 scores 96.4% at top-50 retrieval across roughly 115,000 tokens of conversational history. The next best published result is 94.8% from Mem0.

On BEAM, which tests memory at 100K, 1M, and 10M token scales, M-1 scores 76.9%, 75.0%, and 68.0% respectively, leading at every scale. The competitive gap widens at the hardest tier: M-1 leads by 3.5 points at 100K and by 3.9 points at 10M. When the retrieval problem gets harder, the architectural advantage becomes more pronounced rather than less.

Both results were achieved with Gemini 3 Flash. Every other system on both leaderboards used Gemini 3 Pro, which is four to six times more expensive. M-1 also consumed approximately 20% fewer total tokens per query than the next best system on BEAM.

The full methodology, prompts, and results JSON are published for both evaluations. See the research papers for LongMemEval and BEAM.

Why the model choice matters

A larger model can compensate for weaker retrieval. Given noisy or partially relevant context, a more powerful model is better at extracting the correct answer despite the imprecision. Results achieved with a larger model are, in part, measuring the model's reasoning capacity rather than the memory system's retrieval quality.

M-1 achieving higher scores with a smaller model suggests the retrieval architecture is doing more of the work. That is the right place for it to happen in a production system, because retrieval cost is paid on every query while model cost compounds with context size. A memory system that requires a frontier model to perform well is not production infrastructure. It is a research configuration.

What M-1 does differently

Most memory systems store conversation history or extracted facts and retrieve by semantic similarity. M-1 treats memory as reconstructive rather than as simple lookup, drawing on cognitive science frameworks from Tulving (episodic and semantic memory), Bartlett (reconstructive recall), and Howard and Kahana (temporal context models). These are not decorative citations. They describe the actual design of the retrieval pipeline.

The scoring function combines semantic similarity with lexical precision, temporal salience, importance scoring, and cross-memory coherence. Queries are decomposed into parallel retrieval passes targeting distinct information needs. Retrieved candidates are re-ranked for coherence before being assembled into context. The result is a retrieval system that surfaces what is accurate rather than what is merely similar.

This multi-signal approach is what prevents the failure modes that degrade other memory systems at scale. Semantic collapse is mitigated because similarity is one signal among several rather than the sole retrieval mechanism. Memory drift is prevented because contradictions are resolved at write time. Entity resolution links fragmented references automatically as part of memory creation. Context window overflow is avoided because precise retrieval produces a compact, relevant context rather than requiring the model to sift through noise.

The production origin

M-1 was not built as a research project. It was built because Fabric, a consumer knowledge and workspace product, needed a memory layer that could handle real users storing and retrieving information across months and years of use.

The requirements that came from that experience shaped the architecture from the start. Cross-session persistence, temporal tracking, contradiction resolution, and multi-session synthesis were the original problems, not features added later. A memory system built against a real production workload encounters failure modes that benchmarks and demos do not surface, because those failure modes accumulate over time and scale.

How it compares

The best agent memory platforms page covers the full competitive landscape with benchmark results, feature matrices, and pricing. Individual comparison pages exist for Mem0 and Supermemory.

The honest summary: Mem0 has the largest community and the broadest framework integrations, with 48,000 GitHub stars and an open-source codebase. These are real advantages for teams that weight them. M-1 leads on retrieval accuracy at every evaluated scale, using a cheaper model, with a methodology that is fully published and reproducible. The tradeoff between community ecosystem and benchmark-leading accuracy is a real one, and which matters more depends on the team.

More than memory

Exabase is not a memory-only platform. Alongside the Memory API, it provides Deep Search for hybrid sub-document retrieval, Extract for structured extraction from any file or URL, Resources for file storage with automatic indexing, and Bases for isolated multi-tenant workloads. Teams that need memory plus document search plus file storage plus extraction get a unified data layer through one SDK rather than assembling it from multiple vendors.

Getting started

Exabase is model-agnostic and framework-agnostic. It works via REST API, Python and JavaScript SDKs, and MCP support for Claude, Cursor, and Windsurf. There is a free tier and a Scale plan at $149 per month.

See the docs to get started, the examples for working implementations, or the research for the full benchmark methodology and results.

FAQs

What benchmarks does M-1 lead?

M-1 holds state-of-the-art results on both LongMemEval (96.4% at top-50) and BEAM (76.9% at 100K, 75.0% at 1M, 68.0% at 10M tokens). No other memory system holds the top score on both benchmarks at every evaluated scale. See the research papers for LongMemEval and BEAM.

Why does the model choice matter?

A larger model can compensate for weaker retrieval by extracting correct answers from noisy context. M-1 achieved higher scores using Gemini 3 Flash while every competitor used Gemini 3 Pro, which is four to six times more expensive. This suggests the retrieval architecture is doing more of the work, which is the right division of labour for a production system.

What is LongMemEval?

A public benchmark that tests long-term memory in conversational AI systems. It presents roughly 115,000 tokens of conversational history across multiple sessions and tests six memory capabilities: single-session recall, preference tracking, assistant-provided information, multi-session reasoning, temporal reasoning, and knowledge update. See what LongMemEval actually measures.

What is BEAM?

BEAM (Beyond a Million Tokens) tests memory retrieval at 100K, 1M, and 10M token scales. At 10M tokens, the corpus is equivalent to roughly a year of daily conversations with an AI agent. Context windows cannot hold it, so retrieval architecture is the only factor that determines the score. It tests ten memory capabilities including three that no other benchmark covers.

How does M-1 compare to Mem0?

M-1 scores 96.4% on LongMemEval using Gemini 3 Flash. Mem0 scores 94.8% using Gemini 3 Pro. Exabase's evaluation forked Mem0's benchmarking script, removed question-category-specific prompt templates, and used a uniform prompt with a cheaper model. The methodology is fully published. See the full comparison.

What makes M-1's architecture different?

M-1 treats memory as reconstructive rather than as simple lookup, drawing on cognitive science frameworks for episodic and semantic memory. The retrieval pipeline combines semantic similarity, lexical precision, temporal salience, importance scoring, and cross-memory coherence rather than relying on any single signal. See episodic vs semantic memory for AI agents.

Does M-1 handle contradictions automatically?

Yes. When new information contradicts existing memory, M-1 identifies the relationship, determines which is more recent, and updates the memory graph accordingly. The model receives the current state rather than conflicting versions. See what is memory drift.

Is Exabase only a memory platform?

No. Alongside the Memory API, Exabase provides Deep Search, Extract, Resources, Bases, and Workers. Teams that need memory alongside document search, file storage, and extraction get a unified data layer through one SDK.

Can I reproduce the benchmark results?

Yes. The evaluation methodology, prompts, and results JSON are published for both benchmarks. See the LongMemEval paper and the BEAM paper.

How do I get started?

Sign up at exabase.io for a free tier with no credit card required. The docs cover setup and the examples show working implementations.