2026-03-24

• #0002

  #0002Idea signalsHacker News · AI Agent2 months ago

  `Cq`: a shared knowledge layer for coding agents

  50radar

  CqTool — reviewed shared knowledge layer for coding agents

  A local-first KU store lets agents save and reuse verified gotchas across repos. Strong workflow signal for agent memory, but still an early PoC.

  • —Agents write knowledge units in a shared schema, then query them before starting work; this targets repeat mistakes like stale GitHub Action versions.
  • —Default setup is local-only with SQLite plus a Python MCP server, so the pattern is cheap to test without adding hosted infra.
  • —Team mode adds FastAPI, Docker, and browser review with human approval before knowledge becomes searchable by other agents.
  • —The useful wedge is narrower than CLAUDE.md or AGENTS.md: short, verifiable fixes instead of growing prompt-rule files that become noisy.
  • —This is less a finished tool than a product cue: shared agent memory with review, confidence scoring, and privacy controls has room to become real infra.

  Source: blog.mozilla.ai/cq-stack-overflow-for-agents/Read original →
• #0001

  #0001Agents & toolsHacker News · MCP Server2 months ago

  `Cq`, a shared memory layer for coding agents

  50radar

  CqTool — local-first shared memory layer for agents

  Instead of stuffing AGENTS.md, it stores verified gotchas as reusable knowledge units any agent can query first. Local-first design plus optional team review makes it practical to test now, though public-network value is still unproven.

  • —Agents submit structured knowledge units from real failures, then query them before starting similar work to avoid repeating the same mistakes.
  • —Default setup is local-first with FastMCP + SQLite; knowledge stays on-device unless you opt into team sync.
  • —Team mode adds FastAPI and a browser HITL review flow, so shared memory is curated before other agents can consume it.
  • —One concrete win: outdated GitHub Action major versions were captured once, then reused across repos and models after proactive verification.
  • —The interesting bet is the interface: cross-agent memory as a separate layer, not more rules crammed into CLAUDE.md or AGENTS.md.

  Source: blog.mozilla.ai/cq-stack-overflow-for-agents/Read original →