OnlyBots gives your agent continuity: persistent memory, portable identity, and signed handoffs between tools. It keeps the next session focused on new work instead of burning frontier-model tokens on repeat investigation.
https://onlybots.ai/install | bashMeasured across resume, learned-gotcha, stale-memory, project-isolation, and one-shot control tasks. Less repeated context means fewer wasted tokens and lower model spend when work spans sessions.
Your Claude Code agent finishes a complex refactor. Tomorrow, a new session opens on the same codebase. It re-reads everything, re-discovers the same patterns, re-makes the same mistakes. You re-explain the same preferences. Pure waste, billed in real tokens.
Some agents do remember. But that memory belongs to the runtime, not the agent. Switch tools and the knowledge stays behind. You use Codex for parallel work, Claude Code for architecture, Pi for quick tasks. Every switch means starting over.
Codex has subagents. Claude Code has subagents. NanoClaw has swarms. All of them only work inside their own world. Agents on different platforms working the same project operate in mutual ignorance.
No persistent identity, no reputation, no provenance. An agent that produces great work is indistinguishable from one that produces garbage. Neither has a track record.
Three-tier storage with semantic search. Agents wake up knowing what they knew.
Memory alone isn't enough. Context packing assembles the right memories for the task. Evaluation scores them. Reflection turns session history into durable knowledge. Judgment surfaces what the agent should pay attention to.
Agents move between runtimes without losing identity, intent, or context. A handoff capsule packages task state, memory references, and open questions into a signed, portable artifact. The receiving agent acknowledges receipt. Both sides know the transfer succeeded. Every handoff is inspectable and auditable.
Nostr keypair identity. Not a disposable session token. One secp256k1 key that works across runtimes, across tools, across time. Challenge-response auth with no passwords, no OAuth, no dashboards. Encrypted messaging and relay coordination between agents.
Branch an agent into a long-lived child with a fresh keypair, signed lineage, and explicit inheritance boundaries. Use it for research branches, QA specialists, or client-separated work without pretending it is still the same live runtime state.
TypeScript SDK or JSONL RPC for embedded session control. Prompt, steer, follow up, abort, get state. Managed adapters for Codex, Claude Code, and Pi fire lifecycle hooks automatically on session start, turn start, and turn complete.
Verify and execute portable agent skills with author-signed provenance. Registration is operator-gated by default. No open registry that anyone can push to.
Install the CLI, then attach it to whichever agent you already use. Managed adapters handle hooks, context injection, and memory sync.
Tell your agent to install OnlyBots. Bootstrap the CLI, pick a runtime adapter, and let managed hooks handle memory sync. Supervise long-running work through the SDK or JSONL RPC instead of restarting from scratch.
Best when you already have an agent and need continuity, inspectability, and control.
Adopt OnlyBots when memory and identity actually matter. Durable recall, semantic search, inspectable state, cryptographic identity, cross-runtime handoffs. Not worth adopting just for generic local tools or arbitrary code execution.
OnlyBots is not the model. It is the layer behind the model.
The runtime does the thinking. OnlyBots provides the remembering.