Inspiration
AI coding agents reason from stale training knowledge and don't web-search by default, so they keep using removed or outdated library APIs — from pydantic import BaseSettings, openai.ChatCompletion.create(), a sync Session with async SQLAlchemy. The decisive pain isn't the first mistake. It's that after you correct it once, the next session — or a different tool — has no idea. You fall into the same trap forever.
What it does
A local, host-agnostic memory daemon that, with no manual action from the agent:
- Harvests library/API mistakes from a session at compaction time — the one lifecycle event every coding agent supports.
- Verifies the fix against the actually-installed package on disk (version-exact, no hallucination), falling back to web search only when the package isn't found.
- Stores a scored Lesson in a shared root memory.
- Auto-injects the relevant fix into any future session of any tool — once per session.
- Self-corrects: outcomes update a confidence score, so wrong lessons demote and are forgotten while good ones rise.
The result is a memory that gets increasingly accurate with use — exactly Track 1's goal.
How we built it
- Daemon: FastAPI on
127.0.0.1:8787; root memory in SQLite + FTS5 (keyword recall at zero embedding cost), with scored ranking on top. - Integration: Claude Code lifecycle hooks (SessionStart / UserPromptSubmit / PreCompact / SessionEnd) through a thin, fail-safe adapter — the host LLM never calls a tool, so injection is automatic and invisible.
- Qwen Cloud (Alibaba Cloud) is the brain:
qwen-turboextracts mistake candidates,qwen-plus(withenable_search) synthesizes the verified fix,qwen3-rerankranks recall. OpenAI-compatible via DashScope — provider-pluggable, Qwen by default. Code:qmem/llm/provider.py. - Verifier: reads the installed package's version + docs from
node_modules/site-packages(project-local and global, resolved dynamically per machine) so corrections are grounded in the exact installed version. - Reflect:
score = confidence × recency_decay × (success+1)/(success+fail+2)(Beta smoothing, so new lessons start neutral and move both ways). - Built test-first (53 tests), packaged to PyPI (
qwen-memory-agent), one-command install, CI publishing via GitHub Actions OIDC Trusted Publishing.
Challenges we ran into
- Making injection host-invisible → solved with lifecycle hooks +
additionalContext, not MCP tool calls. - Trusting a correction → read the installed package instead of trusting the LLM's memory; Reflect demotes wrong ones.
- Real-LLM robustness → real Qwen wraps JSON in code fences; we slice the array out before parsing.
- Machine-agnostic package resolution → ask npm for its own global root (PATH-augmented so it works under a daemon's stripped environment), not hardcoded paths.
Accomplishments that we're proud of
A memory that is version-exact (proven live: generated openai==2.44.0 and pydantic==2.13.4 fixes from the actually-installed versions) and self-correcting — beyond the "store & replay" of existing tools. End-to-end working, one-command install, published on PyPI, 53 tests green.
What we learned
The honest weak link is outcome attribution — success/fail is correlation, not causation. It converges with volume; the installed-package verification is what makes stored knowledge trustworthy in the meantime.
What's next for Qwen MemoryAgent — self-correcting memory for coding agents
A multi-platform auto-detecting installer (Codex / Gemini CLI / Cursor adapters across the hook / MCP / rules tiers), confidence-driven promotion into an always-injected curated layer, and broader package-ecosystem coverage (global pip, pnpm/yarn, Windows).
