Agents that remember what matters.
Neutron AI turns user preferences, project state, skills, knowledge, and tool lessons into controlled memory that agents can reuse without replaying every old conversation.
25-60%
target context reduction
<100 ms
hot-pack retrieval goal
0 raw
sensitive text in recall output
SpaceX memory demo
Local demo facts, scoped recall
Ask a SpaceX question. I will answer from a compact demo memory pack and show which memories were recalled.
Memory that behaves like product infrastructure.
Every memory write, recall, compaction, and cache decision has a visible boundary, a stable contract, and a clear failure path.
Capture
Agents write memories into explicit tenant and scope boundaries.
Compress
Stable facts, preferences, skills, and project state become compact memory cells.
Recall
Queries retrieve only authorized memories with deletion-aware scoring.
Pack
Reusable context packs give agents fast, consistent working memory.
A control layer for agent context.
Give product teams and developers the same operating model for memory, prompts, skills, knowledge, connectors, and evaluations.
Agent memory
Long-lived facts, preferences, lessons, and task state.
Knowledge bases
Scoped retrieval across approved public and private knowledge.
Prompt operations
Editable prompts, llm.txt files, and skills with controlled rollout.
MCP connectors
Memory tools for agent runtimes and frontier model workflows.
SDKs and CLI
Integration paths for platform teams and local developer workflows.
Evaluation loops
Benchmark speed, size, accuracy, safety, and regression movement.
A small API surface agents can trust.
The platform keeps memory operations explicit: remember, recall, forget, compact, and build a context pack.
{
"tenantId": "workspace_acme",
"scopeIds": ["project:launch", "user:operator"],
"task": "prepare the next agent step",
"tokenBudget": 1800,
"cachePolicy": {
"mode": "prefer_cache",
"includeDynamicRag": true
}
}Built around bounded recall, not blind memory.
Neutron AI is designed for product teams that need durable context with clear controls, observable behavior, and safe deletion semantics.
Tenant and scope validation at every memory boundary
Deletion tombstones that prevent stale memory resurrection
Sensitive raw text excluded from recall responses
Auditable admin controls for features, pricing, countries, and editable content
Practical answers before integration.
Give agents durable working memory.
Start with the demo, then move into platform controls for memory, prompts, skills, and integrations.