add reflection memory

[ngduyanhece]

ByteRover currently supports two types of memory:

Knowledge Memory: Captures and stores long-term conceptual knowledge (e.g. facts, rules, patterns).
Reflection Memory (To Be Implemented): Designed to capture and store reasoning steps taken by vibe coding agents during task execution.
Especially relevant for models that use chain-of-thought prompting or multi-step reasoning, such as OpenAI o3, Claude sonnet-4 thinking, etc.

Problem:

In real-world use, reasoning agents often exhibit overthinking, leading to verbose or suboptimal decision chains. Capturing these intermediate reasoning steps can help:

• Optimize and streamline future executions.

• Provide insight into agent logic.

• Enable reasoning reuse across similar tasks.

Goal:

Implement the Reflection Memory system to persist reasoning steps from agents, similar in structure to Knowledge Memory, but focused on task-specific reasoning paths.

Implementation Plan:

• Reasoning Step Extraction Tooling

• Extend or reuse components from the Knowledge Memory pipeline.
• Analyze agent output to extract logical/programming steps or thought chains.
• Handle agents with explicit thought markup (e.g. Thought: / Action: pairs).
  Evaluation Layer
• Contextually evaluate reasoning steps:
  -Was the final result correct?
  • Were certain steps redundant or incorrect?
  • Can this be pruned, compressed, or labeled?
• What to be stored:
  • Task context (input, goal)
  • Agent reasoning trace
  • Resulting output (if successful)
• Storage & Access Layer
  • We can create a different collection of vector db to store the data