Learn · Articles
Your AI Agent Has Amnesia. Here's How We Fixed It at Manna.
In my last post I showed how a fresh AI agent produces a marketing post for Manna Insect in 15 minutes. Several people asked: how does the agent know what to do?
01 · The problem
It reads the manual first
Short answer: it reads the manual first. Just like a new employee would.
Every AI session starts from zero. No memory of yesterday. No knowledge of your strategy, customers, or pricing. If you just open a chat and start working, you're hiring a brilliant consultant with total amnesia — every morning.
Most people are still at the "open ChatGPT in a browser" stage. But a browser chat can't read your files, can't remember last week's decisions, can't update your project logs. The context is gone every time you close the tab.
02 · The shift
Move to a local agentic tool
The jump that changed everything for us: moving to an agentic AI tool that runs locally, with access to your actual files and folders. That's when AI stops being a search engine and starts being a team member.
At Manna we have 30+ active cases across multiple countries, financial models, an 8-year BSF knowledge base, and a CRM. You can't paste that into a chat window. But an agent that reads your folder structure? It just works.
03 · The system
Three layers we built
Here's what we built — simpler than you'd expect:
Layer 1 — One orientation file. Tells every fresh agent: who we are, how the workspace is organized, what the rules are. Ours includes: "prices always in USD with local currency," "never put cost data in customer documents," and "95% of new BSF industry data is wrong — don't update our knowledge base without my approval."
Layer 2 — Process guides per work type. Separate guides for marketing, sales, financial analysis, customer management, knowledge intake. Agent gets "draft a Kenya quotation" → reads the sales guide → knows which files to read, what format to use, what's confidential. No guessing.
Layer 3 — Local memory per project. Every customer and project folder has a status file: what happened last, what decisions were made, what's pending. New agent reads it and picks up exactly where the previous one left off.
04 · The habit
Document for the next agent
The critical habit: every agent documents its work for the next agent, not for you. What it did, what it concluded, what's pending — in enough detail that the next fresh agent continues without asking a single question. That documentation IS the shared memory.
This is how a 3-person BSF company runs cases on 6 continents without anyone holding all the context in their head. The AI agents are capable. The files make them consistent.
05 · Start here
Create one orientation file
Start here: get an agentic tool on your computer, create one orientation file, and ask the agent to log its work. Everything else grows from there.
What does your AI "memory" look like?
Keep exploring
Related topics and pages
- BSF production course — the operating process these shared files keep consistent
- BSF business course — where documented assumptions shape a case
- Articles library — the written guides this post joins
Read next
- Thoughts from our startup and insect experience — Manna's own startup notes on testing small and standardising the process
- Monitoring and automation in BSF production — what monitoring and automation can make consistent in a BSF operation
- What does Manna Insect actually do? Part 2 — the business validation and mentoring side of the same way of working
- AI-Enhanced Workflow at Manna — the earlier post on running a marketing workflow with a local agent
Next step
Put the files to work
Manna can help organise the material, decisions and records of a BSF case so the next step rests on a shared, written context.