Mining Your Own Agent: A Memory Analytics Layer for Agentic Coding Sessions

This talk introduces a local analytics system for agentic coding sessions. It ingests logs, extracts memories and behaviors using LLMs, and surfaces insights via a UI, enabling deeper understanding of agent actions and failures.

gemma3:12b Ollama nomic-embed-text SQLite Streamlit

Overview

A local analytics system that treats your agentic coding session logs as a dataset — ingesting JSONL transcripts, extracting structured memories, entities, and behavioral patterns via LLM, embedding everything for hybrid retrieval, and surfacing it through a Streamlit UI.

Video

Transcript

Generated 3 days ago

Summary

The speaker discusses the challenge of getting AI models, specifically Claude, to effectively utilize external memory systems like MCP servers for coding. They observe that while evaluations show potential, Claude rarely uses these servers, especially as they become bloated. The speaker found that file system memory, where Claude is directed to use familiar tools like 'Graph' to navigate and retrieve information, works much better, particularly with LLM Wiki-style setups. This approach leverages Claude's existing tool-use capabilities for memory retrieval, contrasting with less successful attempts at vector search or graph queries. The audience learns about a practical memory management strategy for improving AI agent performance.

View full transcript

Speaker 0: AI. Cool. Any questions for Lan? Yes. Speaker 1: We have a question. K. So 1st off, MCP servers for yeah. MCP servers for Memory, for coding, or anything else that you think is gonna help plot code just doesn't. Like, I'm I'm 6 months into building Memory MCP servers, coding MCP servers. Speaker 1: Every time you do the eval, it looks different. Right? So it works if Claude uses the MCP server, but Claude almost never uses the MCP server. Especially if you just keep bloating MCP servers as Speaker 0: well, Speaker 1: City doesn't use it. The 1 thing I've noticed that works is file system memory. You can just tell Claude, hey. This is where all my memory is. Use the tools that you're familiar with, like, Graph and whatnot, and, go through the file systems and find what you need Date 2026. Speaker 1: And this works on top of LLM Wiki style things really well. But if you wanted to, do vector search, if you wanted to do graph queries, almost never happens. Speaker 0: Yeah. I just wanna share that. Yeah. That's great. Speaker 1: Exactly. Speaker 0: Great insight. Yeah. Speaker 2: Yeah. But I feel like it, is good for different things. Right? File system memory are good for certain things. And if you you also have to know to write it Down, like, agents sometimes don't write down things that are important, and you have to explicitly prompt it to write it down. Speaker 2: If it exists, then you can look for it. But if it doesn't, then you're out of luck. Right? But the raw traces are just sitting on your disk, and there's a lot of information in Speaker 1: there with the picture. Yes. So extraction versus retrieval. Speaker 0: Right? Speaker 1: So extraction is completely retrievable. Speaker 2: Mhmm. Any? Speaker 1: The different items. Speaker 0: Anyone else? Cool. Speaker 2: Alright. Speaker 0: Tasks,

Tech stack

gemma3:12b

Google's 12-billion parameter open weights model running locally via Ollama, featuring multimodal vision capabilities and a massive 128K context window.

The gemma3:12b model brings Google's highly efficient 12-billion parameter architecture directly to local workstations via Ollama. Built on 12 trillion tokens of training data, this model excels at complex reasoning, code generation, and multilingual tasks across more than 140 languages. It natively supports multimodal inputs (processing both text and images) and leverages a 128K context window to handle massive documents or long-form conversations. Developers can run this balanced model on standard consumer hardware to achieve close to enterprise-grade AI performance without relying on external APIs.

https://ollama.com/library/gemma3:12b

View projects
Ollama

Deploy and run open-source Large Language Models (LLMs) like Llama 3 and Mistral locally on your machine: achieve private, cost-effective AI via a simple command-line interface.

Ollama is the essential tool for running LLMs locally: consider it the Docker for AI models. It packages complex models and dependencies into a single, easy-to-use application for macOS, Linux, and Windows systems. You get immediate access to models like Gemma 2 and DeepSeek-R1 via a straightforward CLI or REST API. This local-first approach guarantees data privacy and security, eliminating cloud dependency and high API costs. Ollama also optimizes performance on consumer hardware using techniques like quantization, ensuring efficient execution even on standard desktops.

https://ollama.com

View projects
nomic-embed-text

An open-source, high-performance text embedding model featuring an 8192-token context window and fully auditable training data.

Nomic AI developed nomic-embed-text to deliver a highly efficient, open-source alternative to proprietary models. Operating with an expansive 8192-token context window, it outperforms OpenAI's text-embedding-ada-002 and text-embedding-3-small on both short and long-context benchmarks. The model is built on a fully reproducible pipeline (offering complete transparency across model weights, training code, and data curation) and supports specialized task prefixes like search_document and search_query to optimize downstream performance in retrieval-augmented generation (RAG) pipelines.

https://huggingface.co/nomic-ai/nomic-embed-text-v1

View projects
SQLite

SQLite is a C-language library: a self-contained, serverless, zero-configuration SQL database engine embedded directly into the application process.

SQLite is the world's most deployed database engine, functioning as a compact, C-language library (under 900KiB with all features) that eliminates the need for a separate server process. It operates as a serverless, zero-configuration system, storing the entire database (up to 281 terabytes) in a single, cross-platform file. This architecture makes it ideal for countless applications: it is built into all major mobile phones, web browsers, and desktop operating systems. The engine guarantees high reliability, supporting full ACID transactions, and its source code is freely available in the public domain for any use.

https://www.sqlite.org

View projects
Streamlit

Streamlit is the open-source Python library that transforms data scripts into interactive, shareable web applications in minutes (not weeks).

Streamlit is purpose-built for data scientists and ML engineers: It eliminates the need for front-end development knowledge (HTML, CSS, JavaScript). The core API uses pure Python to quickly build powerful data apps, dashboards, and machine learning prototypes. Users leverage simple commands (e.g., `st.slider`, `st.dataframe`) to integrate interactive widgets and display data from libraries like Pandas or Matplotlib. The platform supports rapid iteration, automatically rerunning the script on user interaction, and offers streamlined deployment via the Streamlit Community Cloud.

https://streamlit.io

View projects

Mining Your Own Agent: A Memory Analytics Layer for Agentic Coding Sessions

Related talks