Research lab · memory for long-running AI
Cognitive and persistent memory for AI agents and robots.
We build memory for AI systems that run for years instead of minutes. It strengthens what you keep returning to, lets the rest cool off, and forms its own structure over time.
MetaCognition
Memory for AI systems that run for years instead of minutes.
Memory
- Capture
- Consolidate
- Link
- Recall
Four things a long-running agent needs from its memory.
Three products, built on the same memory substrate.
Each one is in build now — tell us which fits your team and we will bring you into the early group.
Fern
Every meeting. One memory.
A meeting bot that remembers every meeting, in any language your team speaks, and shapes what was said across all of them into the actions actually waiting on you.
Hearth
The company, remembered.
An evolving company brain. It learns what your organisation keeps returning to, lets stale work cool off, and answers with the context behind a decision rather than the document that mentioned it.
Theus
Machines that know who you are.
A multimodal perception engine and SDK for agents and robots that have to be personal — recognising the people, places and routines they work with instead of starting cold each session.
Scroll to lay the plates flat
- (01)Meeting memory
Fern
Every meeting. One memory.
A meeting bot that remembers every meeting, in any language your team speaks, and shapes what was said across all of them into the actions actually waiting on you.
Coming soonRead more - (02)Perception engine
Theus
Machines that know who you are.
A multimodal perception engine and SDK for agents and robots that have to be personal — recognising the people, places and routines they work with instead of starting cold each session.
Coming soonRead more - (03)Company brain
Hearth
The company, remembered.
An evolving company brain. It learns what your organisation keeps returning to, lets stale work cool off, and answers with the context behind a decision rather than the document that mentioned it.
Coming soonRead more
Memory infrastructure for teams building AI that has to last.
Three ways companies work with us today. All of it runs on the same memory substrate, and all of it is grounded in published research rather than prompt tricks.
01 - Memory layer
A long-term memory API your agents can call.
MetaCognition turns text, audio, video, and tool activity into structured, persistent memory. Agents can retrieve context across sessions without carrying entire histories in every prompt.
02 - Architecture
Memory shaped around your domain.
We build memory around your product, your team, and your existing tech stack. No vendor lock-in, so you can change the underlying infrastructure whenever you need to.Different domains need different memory architectures. We adapt the memory layer to your data, workflows, and engineering constraints.
03 - Evaluation
Evidence of what your agent retains.
We evaluate your memory layer every week using real production data. We measure retrieval accuracy, knowledge updates, temporal consistency, latency, and token usage, then compare results over time to identify regressions early.Our pricing is based on data ingested, not retrieval volume. Once information is in your memory layer, you can retrieve it without paying per retrieval token.
Benchmarks
Measured over weeks of real use, not a single benchmark run.
Recall latency — retrieval latency
Theus Benchmark Report · Apr 2026 evaluation| Operation | P50 | P90 |
|---|---|---|
| Active retrieval | ~120ms | ~200ms |
| Full system | ~350ms | ~500ms |
Tokens per query — input, typical query · ~30 out
Apr 2026 evaluation · zero LLM tokens at ingestion1,233
Figure 1 · LoCoMo
LoCoMo accuracy by category
Theus production deployment (theus-prod, full system) — September 2026
Where we think memory is going.
A short film on why memory, not scale, is the next constraint on useful AI.
Published
Two papers accepted for publication
In use with a small number of partner teams.
AI you rely on for years should know you better every day not begin again each session.
That is the memory we are building, quietly, with a small number of teams.