Like many of you I run qwen/llama locally for privacy. The missing piece was memory: every session started from zero, and cloud "memory" features defeat the purpose of going local.

So I built MeshCtx. What it does:

  • Cross-session memory stored on your disk — survives restarts, works with the network fully unplugged
  • Memory inspired by human recall: important stuff decays slowly, routine stuff fades (critical memories decay ~200x slower), background consolidation runs every 5 min
  • FSRS-style spaced repetition decides what gets injected into context, so token usage stays sane (tuning params documented in docs/known-issues)
  • Works with any OpenAI-compatible endpoint; built-in presets for Ollama (qwen3, llama3.3)

Offline quick start (or just pip install meshctx):

ollama pull qwen3 curl -fsSL https://raw.githubusercontent.com/LucyAndLuna2023/meshctx/main/install.sh | bash

It also does multi-machine agent swarms (manager/worker), Telegram/Slack/Discord gateways, and releases ship with SHA256 audit evidence + a 4,096-test suite gate (v3.137.8).

Repo: https://github.com/LucyAndLuna2023/meshctx

Not asking for stars — I genuinely want feedback on the memory design from people who run local models all day. What would you want it to remember that current setups forget?