Use KiloCode memory context inside PR-agent

[martijndormans]

Git provider (optional)

Gitlab

System Info (optional)

No response

Issues details

I am testing out this PR-agent in our Gitlab environment. It works rather well at the moment. But I have the following idea and I would like to get some opinions from other people that maybe have some more experience with PR-agent.

My idea is that I think it would be handy to give the PR-agent the memory context of the KiloCode memory-bank. KiloCode is just like Cline or RooCode and thus has a similiar memory-bank principle (https://kilocode.ai/docs/downloads/memory-bank.md). In order for the PR-agent to use this I would need to combine all .md files in the memory-bank to 1 .md file and add it to extra-instructions. Or use the best-practices feature, but that is premium and I am not sure premium would be worth it just for that.

My question about this idea is, if this is something that would actually improve the PR-agent MR review capabilities?
Maybe someone has an even better idea?

FYI: we are using Gemini pro with Openrouter so I think that the context amount would be more then sufficient to handle these long extra-instructions.

[qodo-merge-for-open-source]

💡 Auto-generated documentation-based answer:

Short answer: Yes, adding project guidance can help, but how you feed it matters.

What PR‑Agent supports today:

• Extra instructions (free, OSS/Qodo Merge): You can paste concise rules into the tool’s extra_instructions for each tool (e.g., improve/review/describe). Keep them short and specific; very long monolithic docs tend to reduce model effectiveness.
  Example for improve:

  [pr_code_suggestions]
  extra_instructions="""\
  (1) Prefer X over Y
  (2) Avoid try/except scaffolding for Z
  (3) Ignore *.toml changes
  """
  

• Best practices (Qodo Merge 💎): First‑class support for best_practices.md (local or hierarchical/global). These get applied automatically by the improve tool and produce labeled “Organization best practice” suggestions. This is the cleanest way to inject a repo/org memory.

Practical guidance for your KiloCode memory-bank:

• Don’t dump the entire memory bank as one huge .md into extra_instructions. The docs warn long files/instructions may be processed poorly and can overwhelm prompts.
• If you’re on OSS: extract the most actionable, short, pattern‑style rules (ideally <800 lines total, but far less is better) and place them in extra_instructions per tool (primarily improve and review). Use bullets, keep unambiguous, include tiny before/after examples sparingly.
• If you can use Qodo Merge 💎: convert your memory bank into best_practices.md (or hierarchical best practices via pr-agent-settings repo). This integrates natively with improve and scales across repos/paths.

Will it improve reviews?

• Yes, when distilled into concise, project‑specific rules. It helps the improve tool generate more relevant suggestions and reduces noise. Overly long, generic, or redundant material likely won’t help and may hurt.

Alternatives/better ideas:

• Use local best_practices.md (💎) or hierarchical best practices (💎) for org‑wide reuse.
• Start small: pilot a short best‑practices/core-ruleset; measure suggestion relevance, then iterate.
• Combine with suggestion tracking/auto best practices (💎) to learn from accepted suggestions over time.

Notes for your setup:

• GitLab is supported for describe/review/improve and best practices (💎).
• Gemini via OpenRouter is supported; still keep inputs concise due to prompt quality trade‑offs.

If you share whether you’re on OSS or Qodo Merge and the size/topics of your memory bank, we can suggest a concrete structure (e.g., which rules go to extra_instructions vs. a best_practices.md).

Relevant Sources:

• https://qodo-merge-docs.qodo.ai//tools/improve#extra-instructions-and-best-practices
• https://qodo-merge-docs.qodo.ai//tools/improve#best-practices
• https://qodo-merge-docs.qodo.ai//tools/improve#--tip-"writing-effective-best-practices-files"
• https://qodo-merge-docs.qodo.ai//tools/improve#local-best-practices
• https://qodo-merge-docs.qodo.ai//tools/improve#global-hierarchical-best-practices
• https://qodo-merge-docs.qodo.ai//usage-guide/configuration_options#the-different-tools-and-sub-tools-used-by-qodo-merge-are-adjustable-via-a-git-configuration-file
• https://qodo-merge-docs.qodo.ai//usage-guide/changing_a_model#google-ai-studio
• https://qodo-merge-docs.qodo.ai//index#features

[martijndormans]

On OSS and the size/topics of the memory bank are:

• size: 2300 lines
• topics: rules for react, .NET, general context as in what features the project contains and architectural decisions

[mrT23]

Gemini Pro is a very good model, and is definitely capable of utilizing long context.
Hence, giving it extra context in 'extra-instructions' field is recommended, even if its quite big.

just make sure its 'logical'. For example, maybe add a short introduction line, something like:
"Here is a memory bank that contains ...
use it to better understand the PR code"
Or something similar.

Also, you might also want to increase the maximal context:
https://github.com/qodo-ai/pr-agent/blob/main/pr_agent/settings/configuration.toml#L33
https://github.com/qodo-ai/pr-agent/blob/main/pr_agent/settings/configuration.toml#L134

[martijndormans]

Sounds logical, thanks for the tips I will update this issue when our experiments have been completed.