MemTensor/memmy-agent

The core bet here is that your AI context should outlive any single tool. Memmy positions itself not as another chat assistant but as a shared memory layer that sits beneath Cursor, Claude Code, Codex, and similar coding agents simultaneously. When you switch tools mid-project, the goals, constraints, and dead ends you accumulated don't vanish — they persist in a local memory service running at a fixed address, and every agent that knows to ask gets the same picture of you.

The mechanism is worth understanding concretely. A local memory service (SQLite-backed, default port 18960) stores structured long-term context. An "Agent Source" adapter reads the conversation history of whichever external agent you've been using and ingests it. The README claims this onboarding takes minutes and produces a "First Meeting Report" summarizing months of accumulated project context. Whether that distillation is actually useful depends entirely on the quality of the underlying MemOS retrieval engine, which the README describes as hybrid but doesn't detail further.

The architecture has three distinct entry points sharing one runtime: a desktop Electron app, a CLI/TUI, and an OpenAI-compatible API on port 18990. All three read and write the same memory store. That consistency is the actual product — not any individual interface. The managed Chromium browser tooling is a nice touch: it's bundled so agents never download a browser at task time, and when it's unavailable, other capabilities continue rather than failing hard.

The local-first stance is explicit and specific. Memory, config, and app state stay on your machine by default. When the memory service is unreachable, the system surfaces an error rather than fabricating a response — a design choice the README calls out directly as "no hallucinations," meaning it refuses to invent memories it doesn't have. That's a meaningful constraint to build in.

The comparison table against Hermes and OpenClaw is honest about its own limits — the README notes it's based on public positioning, not a benchmark, and invites corrections. The differentiator it claims is cross-agent memory sharing and the ability to absorb external agent history, both of which the competitors mark as unsupported.

The roadmap signals ambition beyond coding: browser activity, local documents, eventually hardware. Team collaboration with privacy-protected agent-to-agent knowledge sharing is listed as planned. These are large surface areas, and the README is appropriately vague about timelines.

For anyone running multiple coding agents across a single project, the friction Memmy targets is real. Re-explaining architectural decisions to a fresh Claude Code session after spending a week in Cursor is genuinely wasteful. Whether a local memory layer solves that or just adds another service to maintain is the open question this repo doesn't yet answer empirically.