Chat history stores what was said. Department Memory stores what was learned. There’s a fundamental difference, and it’s the reason most AI agent deployments plateau after the initial excitement wears off.
The chat history trap
Every AI platform gives you chat history. It’s a log of messages — what the user said, what the agent responded. Scroll back far enough and you can find that one conversation where the customer explained their requirements.
But chat history isn’t knowledge. It’s raw data. It doesn’t tell the agent:
- What the department’s escalation procedures are
- What the brand voice should sound like
- Which products have known issues
- What the approval thresholds are for different actions
These are the things experienced employees know. And when those employees leave, departments lose this institutional memory — the accumulated wisdom about how things actually work, as opposed to how they’re documented.
The AI parallel
AI agents have an even more acute version of this problem. When a new AI agent is created, it starts with:
- A system prompt (usually generic)
- Maybe a knowledge base (documents it can search)
- Zero awareness of how the department operates
This is like hiring a new employee, handing them a stack of PDFs, and saying “figure it out.” It works, barely, but the agent never develops the kind of deep contextual understanding that makes experienced employees valuable.
What Department Memory is
JieGou’s Department Memory is a structured, always-available knowledge layer per department. It’s not a knowledge base you query — it’s context that’s always present in every interaction.
How it differs from a knowledge base:
| Knowledge Base | Department Memory | |
|---|---|---|
| Trigger | Query-triggered (RAG retrieval) | Always-on (injected automatically) |
| Content | Documents, PDFs, web pages | Rules, preferences, procedures, patterns |
| Growth | Manual uploads | Auto-populates from recipes, templates, workflows |
| Purpose | Answer specific questions | Provide operational context |
When a marketing agent drafts a campaign email, it doesn’t need to “search” the knowledge base for brand voice guidelines. Department Memory injects those guidelines into every interaction automatically. The agent writes in the right tone from the first word.
How it works in practice
Marketing department
Department Memory includes: brand voice guidelines, campaign performance history, audience segment definitions, content calendar context, and approval workflows. When a new content agent is created, it immediately writes in the brand voice, references past campaign performance, and knows which content types need approval.
Customer support department
Department Memory includes: escalation procedures, SLA thresholds, product known issues and workarounds, and customer sentiment patterns. A new support agent knows to escalate billing issues to the finance team, that the legacy API has a known timeout bug, and that enterprise customers get priority routing.
Finance department
Department Memory includes: budget thresholds, approval limits, audit requirements, vendor payment terms, and compliance standards. Every financial workflow operates within the correct governance context from the first execution.
The CLAUDE.md parallel
If you’re a developer using Claude Code, you know CLAUDE.md — a project-level context file that tells Claude about your codebase, conventions, and constraints. It transforms Claude from a generic coding assistant into one that understands YOUR project.
Department Memory does the same thing, but for enterprise departments. It transforms AI agents from generic assistants into ones that understand YOUR department — its processes, preferences, terminology, and accumulated wisdom.
How Department Memory auto-populates
You don’t need to manually write Department Memory from scratch. JieGou auto-populates it from three sources:
-
Installed recipes: When you install a Sales department pack with prospect research, cold email, and deal analysis recipes, the system extracts the operational patterns — what inputs are expected, what outputs are generated, what the typical workflow looks like.
-
Active templates: Template configurations encode department knowledge — the prompt patterns that work, the schemas that define department data structures, the model preferences that reflect quality requirements.
-
Workflow executions: As workflows run, the system captures execution patterns — typical processing times, common error patterns, approval rates. This is Workflow Memory, but the department-level patterns are surfaced in Department Memory too.
Getting started
Department Memory is enabled automatically when you set up departments in JieGou. Install a department pack, create a few workflows, and Department Memory begins accumulating. There’s no configuration step — institutional knowledge builds naturally as you use the platform.