Agent scratchpad not passed to LLM

## Summary
Agent does not reliably use prior step context in subsequent reasoning.

## Current behavior
Later steps can ignore earlier outcomes.

## Expected behavior
Prior steps/scratchpad state influences subsequent decisions.

## Reproduction steps
1. Run multi-step task requiring intermediate memory.
2. Inspect step outputs.
3. Observe repeated/uninformed behavior.

## Hints
- Review state and memory flow in `src/agent/executor.py` and `src/agent/memory.py`.
- Include prior outputs when deciding next step.

## Acceptance criteria
- Multi-step tasks show context carry-over.
- Repeated unnecessary tool calls are reduced.
- Returned steps reflect coherent progression.

## Test note
Use task: “get date then compute offset” and assert no duplicate date calls.

Start by tracing state and memory flow in src/agent/executor.py and src/agent/memory.py, then run the multi-step task “get date then compute offset.” Verify that prior outputs influence the next step, unnecessary duplicate date calls are reduced, and returned steps show coherent progression.

Written by the indexing model from the issue text.