We've all been there. You're working with an AI coding assistant, having a great conversation where it learns your project's architecture, your preferred patterns, the quirks of that legacy module. Then you close the session, open a new one, and it's like you've never met. Blank slate. Every. Single. Time.
Sure, you can paste context into the prompt each time, but that's manual, error-prone, and blows up your token budget. Some folks fine-tune, but that's expensive and slow to update. Others rely on RAG pipelines, but those are complex to set up and often fail to surface the most relevant context at the moment you need it.
What I wanted was simple: a memory system that sits next to my agent, learns from every conversation, and automatically feeds back what matters — without me having to patch the agent's code or redesign my workflow.
That's why I've been using and contributing to Memory Sidecar (v3.1.0), an open-source project that does exactly this.
Memory Sidecar runs as a separate process alongside your agent — think Claude Code, Cursor, Hermes, or any AI tool that writes session logs. It watches a shared data directory, picks up new session content, and processes it through a three-tier memory architecture:
When the agent needs context (or at the start of a new session), the sidecar assembles a tiered context injection — recent hot data, relevant warm snippets, and cold graph recalls — all fitted into a prompt that keeps the agent informed without overflowing.
I've tried several memory solutions for agents, and most fall into two camps: "embed it into the agent" (deep modifications, brittle) or "external RAG" (heavy infrastructure, needs constant tuning). Memory Sidecar is neither. It's a sidecar:
After a few days of use, my coding assistant knows my project: it remembers the module I'm refactoring, the test framework I prefer, the issue we were debugging. No more repeating "remember that bug in parser.py
…" at the start of every session. The recall isn't perfect — it's a best-effort retrieval — but it's surprisingly good at surfacing the most relevant pieces, especially for recent work.
One detail I appreciate: the dossier tracking. If I mention a person's name or a project codename frequently, the sidecar starts building a dossier around it. When I later ask the agent about that person or project, it has structured knowledge ready, not just raw text snippets.
If your workflow is single-session (you start fresh each time and don't need persistence) or if you already have a custom RAG pipeline that works well for you, this may be overkill. Also, it's still evolving — the installation script works but the configuration options are not all documented yet (that's on the roadmap). But if you work with an AI assistant over multiple sessions and hate repeating yourself, this is a practical, no-nonsense solution.
Check out the GitHub repo — it's MIT licensed, so you can use it, fork it, adapt it. The README has install instructions and a quick start. There's also an architecture doc if you want to understand the internals.
Give your AI agent a memory. It deserves one.