Show and tell: Dakera as persistent memory for a code review agent

[ferhimedamine]

What we built

We built a code review agent that uses Dakera to persist context across review sessions. The agent stores:

• PR-level summaries (what the PR does, key risks)
• File history (past review decisions on specific files)
• Author patterns (common mistakes per contributor)

After 3 weeks in production, the agent started surfacing relevant historical context automatically — "this file had a similar race condition in PR #47" — without any explicit retrieval logic.

Stack

• LangChain + GPT-4o for review generation
• Dakera for persistent memory (Docker Compose, single node)
• GitHub Actions webhook to trigger reviews

Setup

Docker Compose took 5 minutes. The Dakera Python SDK made the integration straightforward:

from dakera import DakeraClient
client = DakeraClient(base_url="http://localhost:3000", api_key="dk-mykey")

# Store a review decision
client.store_memory("code-review-agent", {
    "content": f"PR #{pr_number}: {summary}. Key risk: {risk}",
    "importance": 0.85,
    "tags": ["pr-review", repo_name, author]
})

# Recall relevant past reviews before starting a new one
past = client.recall_memories("code-review-agent", query=f"reviews for {filename}", top_k=5)

Results

• Review time cut by ~30% (less re-reading of context)
• False positive rate dropped because the agent remembered previous false-positive patterns
• Zero infrastructure overhead — Dakera runs alongside our existing Docker stack

Happy to answer questions about the setup. The 88.2% LoCoMo benchmark score is what sold us on Dakera over alternatives we evaluated.