AI coding assistants are great in the moment, then wake up tomorrow with the memory of a goldfish. Recallr gives them a diary: a local memory layer that turns not just Copilot CLI & VS Code Chat, even Claude sessions into searchable summaries, activity graphs, and MCP tools. In this 15-minute demo, I’ll show how past chats become reusable engineering context, how assistants can answer from evidence instead of vibes, and how teams can stop re-solving yesterday’s problem with today’s coffee.

This session addresses a growing problem in AI-assisted development: the work is happening in chats, but the memory is not. Decisions, dead ends, diffs, preferences, and lessons get buried across Copilot CLI, VS Code Chat, Claude Desktop, and other sessions. The next day, the assistant starts fresh, the human repeats context, and the team re-solves yesterday’s problem with today’s coffee.

Attendees will learn how to:

1. Turn raw AI coding sessions into reusable engineering knowledge with session discovery, generated summaries, quality ratings, searchable indexes, and links
back to the original conversation.
2. Design AI memory with guardrails: keep data local, expose small MCP tools for search/log/get, and make retrieval visible so people can trust the answer
instead of accepting assistant vibes.
3. Keep memory useful instead of noisy by scoring summaries, purging low-value sessions, capturing concise learnings, and preserving evidence.

I’ll deliver this as a live demo of Recallr. I’ll start with a fresh assistant that cannot remember a previous project decision, then search prior sessions, open the generated summary, briefly show the activity graph, and have the assistant call the MCP-backed recallr tool to answer from evidence. I’ll use screenshots or GIFs as backup, plus a short architecture diagram: session sources -> parsers -> summaries/search index -> MCP and VS Code surfaces.

My perspective is unique because this is a shipped local tool, not a theory slide about future AI memory. Recallr combines a ranked search across using FTS5, dense quantization for embeddings with a semantic sidecar, and comes with a skill installation, and a custom MCP server to get high quality results into your session.

The core lesson is simple: the assistant does not need an even bigger prompt. It needs a librarian.