AI memory: where Mycelora stands
Mycelora is a persistent memory shared between Claude, ChatGPT, Mistral Vibe and Perplexity. It plugs in through an MCP connector (a standard that links an AI to outside tools) and a plugin. This page compares it with the memory built into AI platforms, other shared memories, and memory layers for developers, and notes where each one does better.
Three families of memory
| Family | Examples | Who it is for |
|---|---|---|
| Memory built into one platform | Claude, ChatGPT, Gemini, Le Chat, Copilot | People who use a single AI |
| Memory shared across several AIs | Mycelora, MemoryPlugin, Basic Memory, Supermemory | People who work with several AIs |
| Memory layer for developers | Mem0, Zep, Letta, Cognee, Hindsight, Mastra | Teams building their own agent |
Mycelora belongs to the second family. Like several vendors in the third, it publishes its measurements.
Mycelora vs your platform's built-in memory
| Mycelora | Built-in memory | |
|---|---|---|
| Several AIs | One common memory, read and written by every connected AI | One memory per platform, which another AI cannot query |
| Format | Typed, dated memories: current rule, pitfall, rejected idea, reference point, status | Free-text notes or summaries |
| Handling changes | A replaced or rejected memory is no longer recalled; its history stays available | Per their help pages, memories are added or edited when you ask |
| By project | Spaces, a project brief kept up to date, a handoff between two conversations | Claude and ChatGPT keep a separate memory per project |
| Price | In addition to your AI subscription: Free €0, Pro €10 excl. VAT, Max €30 excl. VAT per month | Included in the platform subscription |
Your platform's memory is included, and it is enough for many uses. Mycelora is for people who switch between AIs depending on the task and work on the same projects for months.
Mycelora vs other shared memories
| Mycelora | MemoryPlugin | Basic Memory | Supermemory | |
|---|---|---|---|---|
| Connection | MCP connector and plugin | Browser extension, MCP, custom GPT, API | MCP | MCP, plugins, API |
| What is kept | Typed memories, project briefs, lasting facts drawn from email and calendar | Memories, chat history, files | Notes as Markdown files | Files, chats, web pages |
| Outdated memories | Replaced at write time within a project, contradictions flagged in the conversation | Updates to outdated facts are suggested, and you approve them | Files you edit yourself | Updated, merged or forgotten by its engine |
| Hosting | Paris, processing in France and Germany | Your machine, or a cloud plan | Cloud, or your own servers | |
| Published scores | BEAM 100K and 1M | LongMemEval, LoCoMo, ConvoMem | ||
| Price | €0, €10 or €30 excl. VAT per month | $89 or $180 per year | Free locally, cloud from $19 per month |
Empty cell: information not found on the vendor's site.
What they do better. MemoryPlugin lists more than 21 supported tools and imports your chat history. Basic Memory keeps your notes as readable files on your machine, free when run locally. MemoryPlugin and Supermemory also update outdated facts. Among memory layers for developers, Zep also tracks how facts change over time, with a temporal graph.
What BEAM measures at 1M tokens
BEAM (ICLR 2026) scores ten memory abilities over long conversations. At the one-million-token tier (tokens are the unit of text for AI models), Mycelora scores 74.5%, ahead of Mem0 (64.1%), with the same judge (gpt-5) and the same grading rubric. The breakdown by ability shows where the difference comes from.
| Ability | Mycelora | Mem0 |
|---|---|---|
| Detecting a contradiction | 84.5% | 35.7% |
| Tracking updates | 80.7% | 65.0% |
| Recognizing that information is missing | 69.6% | 52.5% |
| Putting events back in order | 67.0% | 53.6% |
| Finding a piece of information | 78.5% | 70.0% |
| Following preferences | 87.6% | 88.3% |
| Following instructions | 86.3% | 85.2% |
| Linking several sessions | 67.1% | 65.2% |
| Reasoning over time | 65.4% | 61.8% |
| Summarizing | 57.7% | 63.5% |
The first four rows show gaps of 13 to 49 points: this is Mycelora's core promise: knowing what is no longer true. Mycelora is also better at finding information (+8.5 points). On preferences, instructions, sessions and time, the two are on par. Mem0 summarizes better.
Conditions: 35 conversations, 70 questions per ability, one run, answers by Claude Opus 5.5. Two identical runs can differ by 1 to 5 points on the overall score, and by more per ability. Mem0's figures are those in its documentation, which does not name the answering model. Other vendors publish higher overall scores at 1M, with other protocols: they are listed with their conditions on the Benchmarks page, where our answers and the judge's scores are published.
Our limits
- Five abilities remain under 70%: summarizing, reasoning over time, ordering events, linking several sessions, recognizing that information is missing.
- When information is missing, Mycelora says so, then gives the closest thing it has. BEAM counts this choice as a mistake.
- Replacing an outdated memory is done by a model, within a single project. It can miss a case, so every replacement is visible and can be undone in the dashboard.
- Today a space belongs to one account. Sharing a space between several accounts, organization administration and single sign-on (SSO) are not available yet: they are planned.
- Mycelora cannot be installed on your own machine or servers. Basic Memory runs locally.
- We are currently running BEAM's ten-million-token tier.
Who Mycelora is for
For professionals who work with several AIs on long-running projects: decisions made, options ruled out, lessons learned the hard way, what has changed since.
Another tool will suit you better if:
- you are building your own agent and want a memory API: Mem0, Zep;
- you want your memories as files on your machine: Basic Memory;
- your team needs spaces shared between accounts and single sign-on.
Frequently asked questions
How is it different from the memory in ChatGPT or Claude?
Their memory lives inside one platform. Mycelora is common to your AIs, files each memory by type and by project, and stops recalling what has been replaced.
Mycelora or Mem0?
Mem0 is a memory layer that developers integrate into their applications. Mycelora is a product ready to plug into the AIs you already use. The choice depends on what you do: build an agent, or work with your AIs.
Is Mycelora a RAG system?
No. RAG retrieves passages from your documents. Mycelora retains what gets decided while you work: the current rule, the pitfall, the abandoned idea. Your AI brings the reasoning, your documents hold the information. Mycelora adds the experience.
Where is my data?
In Paris, at Scaleway, an ISO 27001-certified hosting provider. AI processing runs on open models operated in France and Germany, under terms that exclude training on your data.
Can I export my memory?
Yes. Every memory can be read, corrected and deleted in the dashboard, and the whole set can be exported in open formats.
Sources
Help pages of Claude, ChatGPT, Gemini, Le Chat and Copilot; websites of MemoryPlugin, Basic Memory, Supermemory, Mem0 and Zep; Mem0's evaluation documentation; Mycelora's Benchmarks page. Updated: 11 November 2026.