Your AI coding assistant now arrives pre-briefed — without you doing anything

The problem with “bring your own context”

Most AI memory systems put the burden on you. You have to remember to call session_start. You have to pass the right instance_id. You have to hope the AI reads the system prompt carefully enough to use the brain tools before diving into your question.

In practice: people forget. Or they're in the middle of a debugging session and just want an answer, not a setup ritual. The AI gets used without context. The brain sits idle.

We've had 11,000+ downloads of @cachly-dev/mcp-server in under a month. And zero conversions from people who use the server without credentials — because it still works, just without memory. The root cause: no friction at the right moment.

What changed in 0.10.9

We moved session management entirely into the MCP infrastructure layer.

Editor connects
AI reads tools
AI waits for instructions
User asks question
AI answers (no context)
[maybe] AI calls session_start
[maybe] previous context loads

Editor connects
→ ListTools fires
  → session_start automatic
  → git branch + last commit
  → previous session loaded
  → top lessons loaded
First tool call
  → context injected ✓

How the automatic lifecycle works

Editor connects     →  session_start fires (ListTools handler)
                       · reads git branch + last commit as focus
                       · loads previous session summary
                       · loads handoff tasks
                       · loads top lessons by recall count

First tool call     →  session context prepended to response
                       · AI sees: "Last session: deployed API"
                       · AI sees: "Top lesson: express.raw() before json()"

Background          →  codebase indexed (once/24h)

Editor closes       →  session_end fires automatically
                       · git log → auto-generates summary

You do nothing. The Brain manages itself.

The one thing that still needs you

There's one call that can't be automated: learn_from_attempts. The Brain can't observe your terminal. Only you (or your AI assistant) know what happened and what worked. After every fix, deploy, or discovery — your AI calls it automatically, without being asked:

learn_from_attempts(
  topic       = "fix:docker-healthcheck-ipv6",
  outcome     = "success",
  what_worked = "Use 127.0.0.1 not localhost in ClickHouse config",
  severity    = "critical",
  tags        = ["docker", "clickhouse", "ipv6"],
)

VS Code also got offline queuing

cachly-brain 0.7.8 (VS Code extension) now queues lessons locally when the Brain is unreachable — during offline work, network issues, or before setup is complete. Stored in VS Code globalState, auto-synced on activation and every 5 minutes. Nothing is lost anymore.

Getting started

# One command — detects all editors, writes all configs
npx @cachly-dev/mcp-server@latest setup

Restart your editor. From now on, your AI arrives pre-briefed — every session, every machine, no manual steps.