Agent memory
Swarms, fleets and robots, coordinated without a central controller
Software agents and machines stay coordinated through shared memory instead of live connections. A member that was offline catches up when it next checks in.
At a glance
Coordination through shared memory
SAIHM is built for enterprise-class agent teams, and that includes swarms, fleets and robots.
Robots and units that come and go
Robots lose connectivity, and units are added and retired.
One identity, any number of models
One identity can serve any number of models and agent types at once, all on one memory.
A fleet policy every agent obeys
Each agent gets a store policy: where it writes, which stores it may use, what to do when one is out of reach.
SAIHM is built for enterprise-class agent teams, and that includes swarms, fleets and robots.
Robots lose connectivity, and units are added and retired.
Robots lose connectivity, and units are added and retired. A fleet is often a mix of vendors, generations and firmware, and the knowledge worth keeping is the knowledge that should outlive any one unit.
A unit that has mapped an obstacle, or learned that a dock misreads at a particular angle, has produced knowledge every other unit wants. Writing it to shared memory asks nothing of the others, including that they exist yet. A unit commissioned next month reads the same memory on its first shift.
Fleets also cross ownership lines: a logistics operator, a warehouse landlord, a robot vendor under a service contract. Per-item sharing with revocation fits that shape. The vendor’s diagnostics agent reads the fault history it is entitled to and nothing else, and loses that access when the contract ends. The sharing page covers grants, expiry and revocation.
Fleet data replicates to edge caches, to units and to analytics. When something must be removed, destroying the key that decrypts a memory leaves the replicas as unreadable ciphertext.
One identity can serve any number of models and agent types at once, all on one memory.
One identity can serve any number of models and agent types at once, all on one memory. A chat assistant, a voice agent and an email agent can act as one identity, so each knows what the customer told the others. When a better model arrives, switch; nothing migrates.
Drop-in adapters cover LangChain, LlamaIndex, LangGraph, CrewAI and AutoGen, a core client serves any Python app, and an endpoint you can run yourself is available. Runnable demos in six models are on the developers page.
Available today
Each agent gets a store policy: where it writes, which stores it may use, what to do when one is out of reach.
Each agent gets a store policy: where it writes, which stores it may use, what to do when one is out of reach. Your organization’s identity issues it centrally as a signed cell that the agent verifies and cannot change. Scopes decide what each agent may recall.
| Agent | What it does |
|---|---|
| Verifies its signed policy | Writes to the stores the policy names: the local, central or SAIHM store |
| Cannot verify its policy | Reads its local store only and writes nothing shared |
SHM · every paid plan: store policy and scopes
Shared memory is for knowledge, not reflexes.
- Shared memory is for knowledge, not reflexes. Anything on a millisecond budget, such as motion control, collision avoidance or safety interlocks, belongs on the robot.
- It coordinates agents that already know of each other. Finding an unknown counterparty is a different mechanism.
- Two agents that write separately do not thereby agree on an order, so work that depends on strict sequencing needs something that provides it.
Tune each member to its job with recipes and levers. The Fleet recipe gives visibility and consistent behavior across many agents and models (SHM · every paid plan).
- Is there a central controller?
- No. Agents and machines stay coordinated without one, and no controller has to be online for the fleet to function. Your organization can still issue a signed store policy centrally; each agent verifies it and cannot change it.
- What happens to a member that was offline?
- It catches up when it next checks in.
- What if an agent cannot verify its policy?
- It reads its local store only and writes nothing shared.
- Can one identity serve agents on different models?
- Yes. One identity can serve any number of models and agent types at once, all on one memory.
No card and no wallet; you need Node.js 20.19 or later.
No card and no wallet; you need Node.js 20.19 or later. Any MCP host that runs local servers can add SAIHM with one command, beside the memory you already use:
npx -y @saihm/mcp-server-pro free-joinKeep reading: agent sharing, enterprise-class memory, enterprise, memory security, pricing and the Trust Center.