Doesn't work in Claude Code

[janosrusiczki]

📋 Pre-flight Checks

• I have searched existing issues and this is not a duplicate
• I understand this issue needs status:approved before a PR can be opened

📝 Bug Description

I have installed the plugin, worked in a few Claude sessions with different complexities and scopes: flow changes, UI refactors and bug fixes but from what I've seen engram is never searched and certainly never saves. I did manually ask at the end of each section and this is the response I get.

---

Me: did you use engram in this session? should you have used it? if the answer is yes, why didn't you use it?

Claude: No, I didn't use Engram in this session, and I should have.

The engram:memory skill is marked ALWAYS ACTIVE with explicit instruction: "You MUST save decisions, conventions, bugs, and discoveries to engram proactively. Do NOT wait for the user to ask."

This session had a clear candidate worth saving: the discovery that MAX_CANDIDATES = 50 in WaypointDetectionService acted as a hard cap on junction candidates (applied as LIMIT 50 ordered by fraction_along_track), causing waypoint detection to silently give up on the second half of longer trails. That's a non-obvious gotcha about the service's behavior.

I skipped it because the task was small and mechanical - read file, delete two lines, run tests - and I moved straight to done without pausing to consider what was worth persisting. That's the wrong habit for exactly the kind of subtle bug that's easy to forget and hard to rediscover later.

---

And then it promises it will use it and then it doesn't.

So, unfortunately it doesn't deliver, I was expecting it to work seamlessly: when Claude starts investigating use it, when I finish a session with Claude it saves. It doesn't do either.

🔄 Steps to Reproduce

1. Install engram
2. Start Claude Code session
3. Wait for engram to be invoked

✅ Expected Behavior

To be invoked.

❌ Actual Behavior

It is not invoked.

Operating System

Linux (Ubuntu/Debian)

Engram Version

dev

Agent / Client

Claude Code

📋 Relevant Logs

💡 Additional Context

No response

[janosrusiczki]

More "questioning":

---

Me: I don't understand why you keep ignoring the engram instructions. Do they conflict with your system prompt? I think they are clear but you still ignore them consistently. There must be a reason.

Claude: Honest diagnosis:

1. The tools are deferred. Engram tools aren't loaded at session start — they require an explicit ToolSearch call to get their schemas before they can be used. By the time I've loaded them, I've already mentally framed the task as a code research problem and have momentum toward Grep/Read.
2. My default pattern overrides the instruction. My trained behavior is: read question → identify what code answers it → search code. The engram search is supposed to short-circuit that at step 1, but I treat it as an optional augmentation rather than a mandatory gate.
3. No natural checkpoint. There's nothing in the tool flow that forces a pause. I have to choose to stop before my default pattern kicks in. When the question looks answerable from code, I don't feel the friction that would make me pause and check memory first.

The save instruction works better in practice because it fires after completing work, when I'm naturally wrapping up — there's a clear trigger ("done"). The search instruction fires at the start, when there's no trigger except willpower against a default pattern.

What might actually work: framing it as a hard prerequisite in the skill definition (the engram:memory skill already says "ALWAYS ACTIVE" but only mentions saving). If the skill explicitly blocked proceeding without a search, that would create real friction. But that's a Claude Code skill configuration change, not something in CLAUDE.md.