Execution Memory Runtime for AI Agents

Aionis gives long-running AI agents a shorter, cleaner, auditable execution context.

It is a memory Runtime for Agents that need to keep working across sessions, threads, roles, devices, model boundaries, and compaction boundaries without dragging the whole history forward.

Aionis records plans, decisions, outcomes, validation evidence, procedures, blocked alternatives, ordinary facts, and rehydrate pointers. Then it decides what is current, what needs inspection, what should stay blocked, and what must be expanded on demand before compiling the next Agent context.

observe -> guide -> agent action -> feedback -> measure -> snapshot

Start With One Setup Command

The product path is Runtime first, then the integration surface that matches your Agent host: SDK, HTTP, MCP, AIFS, native plugin, multi-agent host contract, or Memory Firewall for external memory.

npx aionis setup

The guided setup installs a local Runtime, writes .env, and configures the embedding provider used for stored-memory recall. Optional AIFS, Zvec ANN, and native hook integrations can be added during setup.

For the shortest real Agent connection path after install, use Connect an Agent. It shows the SDK, HTTP, MCP, and AIFS options and the exact final context each path gives to the Agent.

The product loop is the same in every integration:

observe -> guide -> agent action -> feedback -> measure -> snapshot

For MCP-compatible coding Agents:

claude mcp add --transport stdio --scope project aionis -- \
  npx -y @aionis/mcp@latest \
  --base-url http://127.0.0.1:3001 \
  --scope-from workspace \
  --workspace-id-store user

Then start context-first:

aionis_context -> Agent action -> aionis_record_step -> aionis_flight_recorder

For Claude Code sessions, install the plugin after the Runtime is running:

/plugin marketplace add https://github.com/ostinatocc/aionis-claude-code
/plugin install aionis@aionis-claude-code
/aionis:onboard

What Aionis Gives You

Aionis surrounds an Agent loop with governed execution memory. The Agent host plans, executes tools, validates work, and owns permissions. Aionis records evidence, governs what memory can influence the next turn, and explains every decision.

The Core Difference

Full history answers:

Pass everything to the model.

Recall memory answers:

What memories are related to this query?

Aionis answers:

Which memory state is allowed to affect this action right now?

That difference matters when full history is too large, retrieved memory is noisy, or old execution evidence and stale assumptions remain semantically relevant. Aionis separates use_now, inspect_before_use, do_not_use, and rehydrate so related history becomes governed state before it becomes instruction.

Product Surfaces

Execution Memory is the primary Runtime capability. Memory Firewall is the fastest adoption path when you already have another memory backend. Trace-to-Skill Candidates are the learning surface: Aionis can turn positive continuity and workflow-reuse evidence into reviewable candidate assets through trace -> feedback attribution -> measure -> candidate -> review -> promotion gate, while keeping them out of Agent prompts until normal admission and promotion gates approve later use.

Invalidated-route handling remains part of the safety model: rejected or stale execution evidence stays governed. The headline is state-preserving, execution-ready context at lower context cost than full-history transfer.

Current Evidence

How It Works

1. Your host calls observe when the Agent attempts work, sees tool output, gets verifier feedback, or finishes a handoff.
2. Your host calls guide before the next Agent turn.
3. Aionis compiles governed context: current path, inspect-only evidence, governed alternatives, procedures, and rehydrate pointers.
4. The Agent acts in your host.
5. Your host sends feedback tied to the guide trace.
6. Aionis can measure effect and expose a snapshot for replay.

Start Here