v1.2 · Open SourceYour agents forget. Myco remembers, with provenance.
An open-source memory layer for AI agents that self-hosts on your own Postgres and runs with no API keys. The program writes the facts, not the LLM, so every fact it keeps carries a source, a confidence score, and an audit trail.
Free and open source to self-host · No API keys required to start
One memory · every MCP client
From clone to a connected agent, in minutes.
One docker compose up boots the whole stack, no API keys required to start. Then paste one command into Claude Code, Cursor, Codex, Claude Desktop, Windsurf, or any MCP client.
Starts Postgres 16 + pgvector, the MCP server, and the extraction worker, with a seeded workspace, so it works with zero configuration. Full-text search and ingestion work immediately, no API keys, and with local Ollama embeddings, semantic search is keyless too.
Turn your ChatGPT or Claude history into memory.
Point Myco at your data export and every conversation becomes searchable, provenance-tracked memory, one document per conversation, on your own Postgres.
Branched ChatGPT conversations import the transcript you kept, not the regenerations you threw away.
Content-hash dedup means importing the same export twice never multiplies your memory.
Ask brain_why and each imported fact links to the exact conversation in your export.
This is what it feels like when your AI actually remembers.
These aren’t hypothetical, they’re the verified demos from the README. Run them yourself with the included demo corpus: recall, sharing, and provenance need zero API keys; the graph demo runs fully local with Ollama.
The new session retrieves the stored fact instead of relying on chat history. That's cross-session recall.
Both clients read the same shared memory, because the source of truth is Postgres, not a single chat thread.
brain_why returns the provenance chain, “Supported by 4 mentions across 4 source documents”, not a trust-me summary.
Graph queries surface connected people, documents, and entity-to-entity edges, not flat vector matches. Buildable locally with Ollama, no API key.
Five verified demos ship in the README, every answer traced back to its source document.
Where Myco is indispensable.
Tell us who you are and your industry, and see exactly what becomes possible, in the industries where memory you can audit is a requirement, not a nice to have.
FinTech
RegulatedThe problem: Regulators now demand a replayable audit trail for every AI decision.
Wire your fraud or ops agent to one memory layer with an RLS-enforced, replayable decision trail. Access control is deterministic, not LLM-trusted.
Why Myco: Every decision traces to its source and rule via brain_why, the program writes the facts (no silent hallucination into the record), and it self-hosts so financial data never leaves your VPC.
Stateless agents forget. Black-box memory can't be trusted.
Your agents start from zero every session, so you re-explain the same context forever. The memory that does exist is a black box: it rewrites itself silently, with no way to see where a fact came from or whether it still holds. Worse, recent research shows the damage compounds: agents fed noisy history degrade, and swapping in clean records restores behavior. The failure stays silent until production is already paying the price.
Built for builders serious about AI.
If your agents forget, this is for you. Whether you’re a developer shipping to production, a vibecoder moving fast, a team putting AI in your product, or an agency running agents across clients, Myco is the memory layer underneath.
We hit this wall at our own agency. Then we built the fix.
The Good Guys has been running autonomous AI agent workflows across our client accounts since early 2025. Every agent started from scratch. Agents contradicted each other. We spent an hour every morning assembling context for work that should have been instant.
We built Myco because we needed it. Today, every agent we run, across every client, queries a single, growing, auditable memory. We don't brief agents anymore. We don't repeat context. We don't correct contradictions.
Now it’s open source under Apache-2.0, built in about three months by a growth marketer directing a team of AI coding agents, and opened to every team that wants the same.
And everything we claim ships with a way to check it. The retrieval and QA numbers run from a benchmark that lives in the repo, with the reader, judge, and dataset all named, so you can reproduce them with one command instead of taking our word for it.
From scattered files to structured intelligence.
No complex setup. No custom pipelines. Myco handles the hard parts so your AI can focus on the work.
A knowledge base, not a pile of notes.
Most “AI memory” is a markdown file the model rewrites until it fills with duplicates and confident guesses no one can trace. Myco is structured, deterministic memory on your own Postgres: facts are typed and deduplicated, relationships are indexed, and every one traces to its source.
With self-hosted Myco, your data never has to leave your infrastructure.
Every time your team pastes a client contract into ChatGPT, that document travels to OpenAI’s servers. Every document you feed an AI agent as context crosses someone else’s infrastructure. Most teams have not fully considered what this means.
Why this architecture.
Most AI memory lets the model keep its own notes. At scale, nobody can review them, contradictions pile up, and the model quietly becomes the authority on what is true. Myco puts the rules in charge: the program decides what becomes a fact, the schema is yours, and the LLM works inside your system, not over it.
| Myco Brain | mem0 | LangChain Memory | |
|---|---|---|---|
| Fact extraction | Deterministic write path | LLM-based | LLM-based |
| Hallucinated facts | Constrained out of the write path | Possible | Possible |
| Provenance | First-class via brain_why | Partial | Partial |
| Shared memory | Native Postgres source of truth | Depends on app wiring | Depends on app wiring |
| Data portability | Plain Postgres tables | Vendor / framework shaped | Framework shaped |
| Cross-session recall | Built into the product | Best effort | Best effort |
swipe to compare →
Deep-dive comparisons:
The compounding advantage.
AI tools give you a powerful day-one capability. Myco gives you something better: an advantage that grows every single day.
- ·Every AI session starts from zero
- ·The same context re-pasted into every prompt, every day
- ·Agents contradict each other with no resolution
- ·Knowledge walks out the door when people leave
- ·Raw client documents flowing through external servers
- ·AI stack is exactly as smart as day one
- ✓Every agent starts informed, client history, commitments, context
- ✓Token-budgeted context packs instead of whole documents per prompt
- ✓Single source of truth, agents coordinate, not contradict
- ✓Institutional knowledge grows and persists permanently
- ✓Raw data stays in your own Postgres
- ✓AI stack gets smarter every single day
The teams that build persistent memory into their AI stack today will have a compounding intelligence advantage that is very hard to catch up to in 2-3 years. The teams that don’t will still be pasting context into chat windows.
Everything we planned shipped. Then we shipped more.
The full launch roadmap is live today, plus a stack of things that weren’t even on it. Here’s what is in your hands now, what is in progress, and the bigger bets after that. Direction, not a contract, dates intentionally omitted.
And the headline numbers are measured, not asserted: 73.6% end-to-end QA accuracy on the complete 500-question LongMemEval oracle subset (reader gpt-4o-mini, judge gpt-4o), with keyless recall@5 of 89.2% with local embeddings, 91.6% with the recency reranker. The harness ships in the repo, run it yourself.
- Self-hosted MCP server: 13 tools on your own Postgres 16 + pgvector
- Keyless search: full-text out of the box, semantic via local embeddings
- Provenance on every fact: trace any answer with brain_why
- Compounding confidence: corroboration raises it, contradiction supersedes
- Dynamic schema: your memory grows its own types, gated by you
- Import your ChatGPT and Claude history, deduped and traceable
- Reproducible 73.6% LongMemEval benchmark, shipped in the repo
- Recommended local-model profile for higher-quality keyless graphs
- Role-level playbook enforcement: shared memory read-only at the Postgres role
- Broader benchmark coverage: harder conditions, more categories
- More ingestion sources: first-class connectors
- Managed cloud for teams who would rather not run Postgres (waitlist)
Keyless retrieval, measured, not asserted.
recall@5 asks a simple question: did the top 5 retrieved memories include the one that actually holds the answer? We measured it on the full 500-question LongMemEval longmemeval_s set, distractors and all.
A deterministic blend of relevance and recency, no API key, no network call:
Compare the hybrid and temporal rows. Retrieval is deterministic, we verified these two independent ways with zero drift. (We report QA accuracy on the evidence-only oracle subset, which isolates reasoning, and recall on the full longmemeval_s set, which isolates retrieval against distractors.)
Same engine, zero ops, a glimpse of Myco Cloud.
The open-source core is the product. The managed cloud wraps it in hosted infrastructure and a full dashboard, for teams that would rather not operate Postgres themselves. Not generally available yet; the waitlist gets access first.
Free to self-host. Cloud coming soon.
The full Myco core is open source and free to run on your own infrastructure. Managed cloud hosting is in development and waitlist-only, join the waitlist to get access first.
Open source core · Apache-2.0 license · No credit card ever required for self-hosted
- ✓Deterministic memory core
- ✓All 11 brain_* MCP tools
- ✓Postgres 16 + pgvector on your infra
- ✓One docker compose up to boot
- ✓Apache-2.0 license, own your stack
- ✓No API keys required to start
- ✓Built on the same open-source core
- ✓We run the infrastructure for you
- ✓OAuth connectors for Gmail, Drive, Slack, and more
- ✓Interactive knowledge graph explorer
- ✓Chat with citations, vault & live dashboard
- ✓Waitlist members get access first
Your agents are only as good as what they remember.
Self-hosting is free, open source, and available right now. Want managed hosting instead? Join the cloud waitlist, early access rolls out to the waitlist first.
No spam, ever · Prefer self-hosting? It’s free, grab it on GitHub