Stash
Sign up

Agents that learn from the real world.

Stash is an applied AI lab building continual learning. We help you search, share, and learn from agent traces. Improve your agents automatically rather than relying on handcrafted evals or RL environments.

What we believe

  1. 01

    AI will outproduce humans.

    Soon, agents will produce more work than the humans they work with. Jevon's paradox will come in full force.

  2. 02

    Agent traces are the new oil.

    The tokens that agents produce will become the most valuable data in any organization. They contain the decisions, context, and work output of teams which can be used to train new models.

  3. 03

    Personalized models beat general ones.

    No matter how smart frontier models become, they will not know your specific workflow or style. Personalized models with operational knowledge of your SOPs will outperform.

  4. 04

    Modularity unlocks future scaling.

    The next frontier of scaling will be based on enabling models and agents to build on top of each other rather than handcrafting evals and pipelines.

Where to start

Internal agentsThe agents your team runs

Coding agents that keep what they learn. Import the logs you already have and stop repeating yourself.

See the product →External agentsThe agents your customers use

Memory for production agents, documented end to end: capture, refine, recall, isolate.

For external agents →

Research

Memory as a world model, not retrieval

Pure search is not sufficient for good memory. We argue for the need to build “world models” where agents develop internal representations of their environment.

Aug 2026Blast radius

Information can expire, be outdated, or only apply in specific situations. Humans are great at these judgement calls around scope but agents are currently lacking.

Jul 2026Stability under accumulation

Memory stores degrade over time due to slop, context rot, and entropy. We measure this problem using a technique inspired by PCR.

Jun 2026