Mnemo Cortex

Not storage. Not memory. Not logs.

Continuity of intelligence across agents, time, and tasks.

Other memory systems are filing cabinets. Mnemo's memory works while you sleep.

Multiple agents. Separate execution environments. One unified memory system.

AI stops resetting. Knowledge compounds. Systems evolve instead of restart.

Not keyword lookup. Not file search. Meaning-based retrieval.

Returns: winning creatives, audience signals, decisions. Not just logs — insight extraction.

One agent executes. One builds. One architects.

All of them: read each other. Learn from each other. Evolve together.

Every night at 3 AM, Mnemo reads every agent's memories from the day. An LLM synthesizes them into a single brief: what was built, what was decided, what's blocked, what each agent should know about the others' work.

Every agent wakes up caught up. No manual relay. No "hey, go tell the other one what you just did." It just happens.

Cost: one cheap LLM call per night. The dream brief is searchable like any other memory — agents can recall dreams from weeks ago. Cross-agent awareness that compounds over time.

You work from two desks. Both machines run Mnemo, and they drift — the decision you saved at one desk doesn't exist at the other. The usual fixes put your AI's working memory on somebody else's wire, or need infrastructure you don't want to run.

The Cortex Stick is a USB courier between two full Mnemo installs. Nothing runs on the stick; it carries the delta. Plug in, sync, pull it out, carry it, plug it in — the other machine catches up. No cloud, no VPN, no account.

“The file about X” is a memory problem too. The Librarian is a single index over your whole workspace — filenames, paths, and the first chunk of every file's content, with text pulled out of PDFs and Word documents — so an agent turns a fuzzy description into a real path in milliseconds.

Our own deployment covers ~107,000 files. A full rebuild takes about 17 seconds; the nightly refresh takes about 2. Secrets are excluded from the index entirely — keys, .env files and credentials never get indexed, so asking the Librarian a question can't surface them.

We ran a compiled wiki in production for months, inspired by Andrej Karpathy's LLM Wiki pattern and Nate B Jones's analysis of write-time vs query-time memory. The lesson was blunt: compiling knowledge into pages is expensive to keep fresh, while indexing everything and finding it on demand is cheap and never stale. So we kept the pages — they're still searchable — and stopped recompiling them nightly. Shipping the replacement mattered more than defending the original.

Multi-agent message bus with full delivery confirmation. Originally lived inside Mnemo Cortex; now ships as its own product.

A reference-grade safety layer for developers building agent systems. Captures how a user works — tone, density, formality, workflow choices — so agents can adapt to them instead of forcing the user to adapt.

Observations become candidates. Candidates get reviewed. Only stable claims promote into the user's profile. Nothing auto-lands.

Five MCP tools: passport_get_user_context, passport_observe_behavior, passport_list_pending_observations, passport_promote_observation, passport_forget_or_override. Reference integration via stdio MCP. The hosted HTTP wrapper for browser-based AIs (claude.ai custom connectors, etc.) is a future release — today's release is for developers who wire MCP subprocesses into their own agent stacks.

Mnemo Cortex captures conversation memory automatically — what agents said, what happened, what was decided. mnemo-plan is the opposite: the stuff you write and curate. Project specs. Active task lists. Decision logs. Architecture docs. Anything an agent needs to know before a conversation starts.

It's a folder of markdown files in Git. Any LLM that can call the Mnemo MCP tools read_brain_file / write_brain_file / list_brain_files can read and edit them. Not Claude-specific.

The starter template ships with project.md, active.md, stack.md, decisions.md, plus optional people.md and incidents.md. Each file has comments explaining what goes there. Fork it, fill it in, restart your agent.

The split: Mnemo Cortex = automatic conversation memory. mnemo-plan = manual project pad. Both ride the same MCP bridge; they auto-enable based on whether BRAIN_DIR is set on disk.

Mnemo Cortex is no longer a memory store. It's a memory architecture.

We adopted the best ideas in the air, credited them openly, and built on top.

Mnemo Cortex is not a wrapper, a bridge, or a plugin for someone else's memory cloud. It's the whole system — storage, recall, and overnight maintenance, running on hardware you own.

Most memory products make you choose: one shared store for everything. Mnemo lets you architect for your actual privacy and separation needs.

Most systems: summarize = discard detail.

Mnemo-Cortex: compress memory, keep traceability.

Go from high-level summary down to the exact moment. ~80% compression. Full fidelity preserved.

SQLite. Local compute. Any LLM for compaction — Ollama for $0, or any API provider you choose.

Over time, Mnemo-Cortex builds: what worked, what failed, what changed. Without re-training.

This is what current AI tools are missing.

They don't blend. They specialize. That's closer to a real company team than a chatbot.

During Claude Fable 5's brief availability, we pointed it at the memory layer it runs on and asked it to make it better. The result shipped as Mnemo's v4.1 "Fable pass."

Mnemo-Cortex stores: ad performance, creative wins, audience behavior, product trends.

Over time, it becomes: a brain for the business. Not a tool.

That's a different category.

You didn't just build something useful. You built infrastructure. And infrastructure is where the leverage is, the lock-in is, the money is.

OpenClaw 2026.4.10 shipped a native Active Memory plugin. Some people have asked whether it replaces Mnemo Cortex. Short answer: no — they solve different problems.

Here's the difference, based on testing both on our sandbox agent.

Built in Half Moon Bay, California by Guy Hutchins and a team of AI agents running on the memory system you're looking at.

AI agents without memory repeat expensive work. Every session starts from zero — re-reading files, re-analyzing context, re-discovering what worked yesterday. That's your API bill paying for the same thinking twice.

Mnemo Cortex remembers what worked, what failed, and what changed. Your agents stop recomputing and start building on yesterday's knowledge. Fewer tokens burned. Fewer expensive model calls wasted on context recovery.

Mnemo Cortex makes every call count — it gives your smart models full context on the first try, so there are no expensive retry loops and no tokens wasted rediscovering what the agent already knew. Your agents stop paying to re-learn yesterday.