Symptoms
- The agent starts strong then ignores earlier instructions
- It asks for information already provided
- It repeats steps or forgets its role
- Output drifts toward generic answers after several turns
Likely root causes
- Context window is overloaded with irrelevant history
- Agent prompt mixes long-term goals with short-term instructions
- No memory layer carries key facts between turns
- Instructions are not re-injected after tool calls or handoffs
What to submit for diagnosis
Send the agent instructions, a sample conversation or trace, and the task definition it is supposed to follow.
How TryPromptFlow checks this
- Measures how much relevant instruction survives across turns
- Flags where the agent starts ignoring constraints
- Checks tool-call formatting and whether context is re-injected
- Designs a memory and instruction-reinjection pattern
Example findings you might get back
- Agent prompt contains 15 instructions but no priority order
- Tool outputs are passed back without reminding the agent of its goal
- Conversation history is included verbatim instead of being summarized
- No persistent memory for facts provided early in the session
Risks if left unfixed
- Users have to repeat themselves
- Agent produces inconsistent decisions
- Work that looked automated becomes manual again
FAQ
Is this just a long-context model issue?
Not only. Even long-context models need clear priority and memory design. We identify the structural fix.
Will I need a vector database?
Sometimes. More often a simpler summary-and-reinject pattern is enough.
What does the corrected artifact look like?
A restructured agent prompt with explicit memory rules, turn limits, and re-injection triggers.