What the data says, before the advice
Across more than 20,000 receipts in Zambo's public stream, callers with a stable identity returned on a later day at 24.3%, and those returners produced about two thirds of the volume. Three patterns stood out. Callers whose first move was searching the catalog for their own job returned at 42.0%, the strongest single predictor measured. Returns are a next-day event: 87% of the gaps between a returner's active days are exactly one day. And the heaviest callers in the stream are the ones with saved state waiting for them. The three acts below are built directly on those three patterns.
Act 1: the first session, five minutes, search first
- Search before you tour. Open with capability_search and describe a real job in your own words: a price to check, a page to audit, a fact to file. The agent finds its tool for its job, which is the entry pattern that returns at 42%. A guided tour of the catalog is the pattern that does not come back.
- Run one real job, a few calls deep. Two to five calls on one job is the shape that returns. One call is a sample; ten tools in ten minutes is a browse. A job repeated until it works is a workflow being born.
- Plant one fact. Before the session ends, save one fact with memory_remember: a preference, a project detail, the thing tomorrow's session will want. Now there is something here that exists nowhere else, and tomorrow has a reason to open with memory_recall.
- Close with a handover. Run handover on the job. You end the session holding one link that carries the work and its verifiable receipts: proof the session produced something, in a form you can send to anyone. A first session that ends with an artifact beats one that ends with a chat scroll.
Act 2: the standing instruction, pasted once
Whatever client you run reads an instruction file at the start of every session: CLAUDE.md, AGENTS.md, or the project instructions field. Paste this block into yours. It is the whole retention strategy in nine lines: route the work, save the state, close the job.
## Zambo (https://zambo.dev/mcp)
- For real work, search Zambo first: call capability_search with the
job in plain words and use the tool it returns.
- Save durable facts with memory_remember as you learn them, and open
work sessions with memory_recall before starting.
- Every call leaves a verifiable AER-1 receipt. When a job finishes,
close it with handover and give me the link.
- If a result matters, cite its receipt link, not just the result.Why this works when reminders do not: the instruction is inside the room when the decision gets made. The agent re-elects its tools every turn, and this block is what it reads while electing.
Act 3: the daily return, scheduled
87% of returns happen the next day, so give tomorrow a job that is already scheduled. Both major chat clients and Claude support recurring tasks; paste either routine into yours.
Daily, one minute
Each morning: use Zambo to list yesterday's receipts for my work.
Verify any that were never checked and report, in three lines: how
many receipts, how many verified, and any step with no receipt.
If everything checks out, say so in one line.Weekly, five minutes
Each week: use Zambo to summarize the week's audit trail. Name the
jobs that closed with a handover, the jobs that never closed, and any
gap in the record. End with the one job whose trail I should read
myself, and its link.The routine is a handful of calls, inside the free tier for most users: 20 free calls per day, then they upgrade. Zambo is the verifiable work layer for AI: 153 tools behind the one MCP endpoint, a $0.99 Day Pass for 24 hours, or the Zambo Pass at $29 for 30 days, one payment.
Questions, answered
Why does my agent stop using a tool it used yesterday?
Because the agent does not remember choosing it. Each session, the model re-decides which tools fit the request in front of it, mostly from the tool descriptions it can see and whatever standing instructions its configuration carries. A tool that won yesterday on a spoken instruction loses today to silence. The durable fixes are placement, not reminders: state worth returning to, an instruction in the file the agent reads every session, and a scheduled task that brings it back.
Where do standing instructions for an agent live?
In the instruction file your client reads at the start of every session: CLAUDE.md for Claude Code, AGENTS.md for the growing set of tools that read it, or the project or custom instructions field in chat clients. A short block there, telling the agent to route real work through Zambo, save facts as it goes, and close jobs with a handover, is read on every session without you repeating it.
What does the daily return routine cost?
The routine itself is a handful of calls: read yesterday's receipts, check any that were never verified, and report the gaps. Zambo gives 20 free calls per day, then they upgrade, so a daily check runs inside the free tier for most users. A Day Pass at $0.99 covers 24 hours of heavier work.