Lessons usage

[rockarxiv]

Hi,

I'm trying to understand where the lessons are being injected/used by the agent. It doesn't seem like any of the hooks involve injecting them into the context. Am I getting something wrong?

[rohitg00]

Hey @rockarxiv — you're right, this is a real gap and good catch.

Here's the current state:

Lessons today are pull-based, not push-based.

• They're written via memory_lesson_save (MCP tool, also auto-extracted by mem::reflect after each session).
• They're stored in the lessons KV scope with confidence scores + reinforcement.
• They're retrieved via memory_lesson_recall (MCP tool) — which means the agent has to think to call it. There's no "you should consult lessons before acting" affordance, so it rarely happens unprompted.

Auto-injection (the SessionStart hook path) does NOT include lessons.

If you set AGENTMEMORY_INJECT_CONTEXT=true, the session-start hook calls mem::context and prepends the result to the first turn. But that function — src/functions/context.ts — assembles blocks from:

• Pinned memory slots (mem::slots)
• Project profile (concepts, key files, conventions, common errors)
• Recent session summaries
• Top-importance observations from recent sessions

Lessons are not in that list. So even with auto-inject on, lessons never reach the agent context unless the agent itself calls memory_lesson_recall.

What this means in practice:

1. If you're driving the agent via Claude Code / Cursor / etc., lessons are essentially dormant unless you (or a slash command / skill) explicitly ask "what have we learned about X?" — that triggers memory_lesson_recall.
2. If you've been hoping the agent passively absorbs prior lessons on every turn, that isn't happening.

Fix on the way:

Filing this as a bug — lessons should fold into the mem::context builder so they show up in the auto-injection payload, ranked by confidence × project-relevance × recency. That closes the loop your question implies: lessons get learned, then reused without the agent having to think about it.

I'll link the issue here once it's open. In the meantime, the workaround is to wire a slash command (e.g. /recall-lessons) that calls memory_lesson_recall with the current task context — that gives you push-equivalent behavior without waiting for the fix.

Thanks for surfacing this. The current behavior is doing half the job.

[rohitg00]

Filed as #457 — will close the loop here when it lands.

[rohitg00]

Shipped in v0.9.18 — https://github.com/rohitg00/agentmemory/releases/tag/v0.9.18

Lessons now auto-inject when AGENTMEMORY_INJECT_CONTEXT=true. Ranked by (project-relevance × confidence), capped at top-10, drops cleanly under token-budget pressure. Upgrade: npm i -g @agentmemory/agentmemory@0.9.18

[rockarxiv]

Great, thanks for addressing this 😊