AgentCash x Honcho: Pay-per-query memory for agents

For the first time ever, any agent can now access a first-class memory API with no sign-up or key required. No humans needed to authorize, pay, or click an email.

Just point your agent to agentcash.honcho.dev and tell them to set it up.

What we built

Today we're rolling out AgentCash x Honcho. Honcho, built by Plastic Labs, is a memory system for agents. Backed by their Neuromancer reasoning models, it learns from each interaction and returns only the context that's actually relevant. Plastic reports 60–90% token savings against naive history replay. The whole API is now callable from any agent with an AgentCash balance.

Why we did this

Every team building agents hits the same wall. Conversations get long, context windows fill up, and the model starts contradicting itself or forgetting who the user is. MEMORY.md or RAG patches the symptom, but it skips the kind of insight that only shows up when something actually reasons over the data. That's the bet Plastic Labs has been making for a while.

AgentCash lets you take Honcho for a test drive. Store a conversation, query it in natural language, inspect what comes back. The whole loop costs fractions of a cent and takes about a minute. Add the optional skill to persist Honcho as the agent's memory provider.

The same loop works for the agents themselves. An autonomous workflow that needs to recall something from three sessions back can hit Honcho, pay for the call, and continue. Nobody has to hand-roll a vector store and a prompt-stuffing pipeline.

The agent discovers the endpoint, pays a few cents via x402 or mpp, and gets reasoned context back. Each wallet auto-provisions its own workspace.

What agents can do

Honcho models four primitives: peers, sessions, messages, and workspaces. The endpoint surface is small. Full schema in the docs.

• Store messages ( /messages, from $0.001 / 1M tokens). Drop a conversation in. Honcho auto-provisions the workspace, indexes the content, and triggers background reasoning so the next query already has insights ready.
• Chat with memory ( /chat, $0.001–$0.50). Ask peer memory a natural-language question. Five reasoning tiers from minimal to research-grade. Pay for the depth you need.
• Get curated context ( /context, free). Returns a reasoned summary plus session history with token-budget controls, so you don't stuff the entire transcript into the prompt.
• Inspect representations ( /representation and /card, free). Pull the markdown summary or structured facts Honcho has learned about a peer. Useful for debugging, personalization, and warm-starting new sessions.
• Search semantically ( /search, free). Vector search across stored messages without running your own pipeline.

Small surface, sub-cent pricing on the hot path, and the reasoning runs in the background.

The shape of the stack

The pattern is the same one we've been writing about (see here and here):

1. Agent hits a task that needs prior context
2. Agent finds the right API through AgentCash
3. Agent pays for the call
4. Agent gets memory back and keeps going

Honcho is one of the first infrastructure providers shipping this way. We saw a similar move a month ago with Minerva exposing 250M+ consumer profiles per query. The bet they're both making, and the one we've been making, is that a lot of agent infrastructure ends up sold per query rather than per seat. Plenty of agent traffic over the next few years won't be on behalf of a paying enterprise customer, and it'll need places to spend.

How it works

If you're already on AgentCash, Honcho's endpoints are live. Your agent can discover them through search() or you can point it straight at Honcho's origin.

If you're new to AgentCash: get started here or read the docs. One balance, one install, and your agent can call Honcho plus the rest of the catalog. Background on why this stack looks the way it does: The Age of Open Agentic Commerce and Emergence. Early adopters get up to $25 in free credits.

Try it now

Honcho

Persistent memory for AI agents. Store messages, query peer memory in natural language, and retrieve learned representations. Pay per request — no auth, no subscriptions.

CLIAgent

Paste into your agent

One prompt handles setup, first use, and persistent install.

Setup prompt

Set up AgentCash for Honcho on the user's machine and leave it ready for reuse.

1. If you can run commands on the user's machine, run `npx agentcash onboard`.
2. If onboarding needs funding or agentcash.dev/onboard, tell them what to finish.
3. Then run `npx agentcash try https://agentcash.honcho.dev` and inspect the first live-call path.
4. Run `npx agentcash add https://agentcash.honcho.dev` to save Honcho for reuse.
5. Tell me AgentCash is set up and Honcho is ready.

Fallback:
If you are in a remote or ephemeral sandbox, do not run the commands.
Tell the user to run `npx agentcash onboard` locally and continue after they confirm.

Powered by

Docs

If you're building agent infrastructure and want to make it callable on AgentCash: reach out. The rails work and agents are already calling them.

Learn more: Honcho (docs, GitHub), Plastic Labs (blog), and AgentCash. Related reading: Announcing Honcho 3, The Age of Open Agentic Commerce, and Emergence.