Show and Tell: Building persistent memory for Python AI agents with Dakera

[ferhimedamine]

What I built

A pattern for giving Python AI agents memory that survives process restarts — using Dakera as the memory backend with hybrid BM25 + vector retrieval.

from dakera import DakeraClient

client = DakeraClient("http://localhost:3000", api_key="dk-your-key")

# Store what your agent learned
client.store_memory(
    "research-agent",
    "User prefers concise, code-first answers with Python examples.",
    importance=0.85,
    tags=["preferences", "style"]
)

# Recall on the next session
response = client.recall(
    "research-agent",
    query="how should I respond?",
    top_k=5
)
for m in response.memories:
    print(f"[{m.score:.2f}] {m.content}")

Why this pattern works

• Semantic recall — finds relevant context even when phrased differently from how it was stored
• Importance decay — high-importance memories (0.8+) persist longer than routine observations
• Zero external deps — one binary, on-device ONNX embeddings, no OpenAI key needed

Resources

• Getting started guide
• Python SDK reference
• Live playground — try it without installing anything

What are you building with persistent agent memory? Share your patterns below!