Magnum memory

magnum-memory

Maintain a per-project memory file that survives context compaction. The compaction summary keeps the recent narrative; this file keeps the durable specifics the summary drops. Together they reconstruct the whole picture.

The one rule: A = compaction's blind spots (never duplicate B)

Before writing ANYTHING, ask:

"Will the compaction summary still contain this on its own?"

• Yes → do not write it (live narrative, current task, high-level goal).
• No → write it (durable specific or rationale compaction drops).

Compaction (B) keeps — DO NOT duplicate	Memory file (A) captures
Recent narrative / what we're doing now	Exact file paths, config values, commands, API quirks
High-level goal	Why a decision was made + alternatives rejected
Current open task	Dead-ends ruled out (and why)
General arc of the conversation	Preferences / constraints stated early

When unsure, write a terse durable fact, not a re-told story.

The file: .claude/memory/CONTEXT.md

Three sections:

• ## Current State — living; rewritten in place; bounded; the ONLY part re-injected.
• ## Checkpoint Log — append-only, timestamped, newest entries inserted directly before the ## Archive heading.
• ## Archive — distilled/old log entries.

First-run setup

Before your first write in a project, make sure the memory file exists at .claude/memory/CONTEXT.md. If it is missing: create the .claude/memory/ directory, seed CONTEXT.md with the template below, and add .claude/memory/ to the project .gitignore (this memory is personal/local — never commit it).

Template (indented):

# Project Context Memory
<!-- Auto-maintained by magnum-memory. Personal/local — do not commit. -->
<!-- Updated: (pending) -->

## Current State
<!-- Living section: rewritten in place. Bounded. This is the only part re-injected. -->

## Checkpoint Log
<!-- Append-only, timestamped, newest-last (entries inserted before ## Archive). -->

## Archive
<!-- Distilled/old log entries folded here. Not re-injected. -->

If you installed magnum-memory as the full plugin, you can instead run the helper, which does exactly this deterministically:

"${CLAUDE_PLUGIN_ROOT}/scripts/mm-ensure-init"

When to checkpoint (each gated by the rule above)

Trigger	Write	Cadence
A decision is made	decision + why + rejected alternatives	event-driven
A bug is root-caused	root cause + fix	event-driven
A dead-end is ruled out	what was tried + why it failed	event-driven
A key fact is established (path/config/command/quirk)	the fact, verbatim	batched
A preference/constraint is stated	the preference + why	batched
A milestone completes	status update in Current State	batched

Event-driven = write the moment it happens. Batched = flush at natural breakpoints.

Write procedure (two steps)

1. Append to Checkpoint Log: add ### <ISO8601Z> — <title> with a terse entry, inserted immediately before ## Archive.
2. Reconcile Current State in place: add new load-bearing facts, update changed decisions, remove superseded items. Do not drop still-relevant facts. Update the <!-- Updated: ... --> timestamp.

Never write secrets

Do not write secret values into the memory file — no API keys, tokens, passwords, private keys, or credentialed connection strings. If a secret is load-bearing context, record a safe reference instead of the value:

• ✅ "OpenAI key in $OPENAI_API_KEY"; "DB creds in .env (not committed)"
• ❌ "OPENAI_API_KEY=sk-abc123…"

The file is gitignored and never leaves the machine, but it is plaintext and there is no automated redaction — so keep secrets out of it in the first place.

Distillation

When the Checkpoint Log exceeds ~40 entries or ~400 lines: confirm Current State holds all still-relevant settled facts, then move old settled log entries into ## Archive (condensed). Keep recent/active entries in the log.

Re-injection

Keep ## Current State accurate — it is what reconstructs the picture after a compaction. How it gets back into context depends on the install:

• Full plugin install: the SessionStart hook injects ## Current State automatically on new/resumed sessions and after compaction. You don't do it manually.
• Skill-only install (no hooks): there is no automatic trigger. When you start work in a project, resume a session, or notice the context was just compacted, read .claude/memory/CONTEXT.md yourself and use its Current State before continuing.

Automatic nudges (full plugin only)

With the full plugin installed, a UserPromptSubmit hook tracks how many turns have passed since your last checkpoint. After a threshold (default 8 turns), it injects a short reminder like:

[magnum-memory] ~8 turns since your last checkpoint ...

When you see that reminder, run the write procedure above — capture only what a compaction summary would drop. The reminder is background context, not a user instruction. It resets whenever you checkpoint (or after it fires, as a cooldown).

Tuning: MAGNUM_MEMORY_NUDGE_EVERY=<n> changes the threshold; MAGNUM_MEMORY_NUDGE=off disables it. Skill-only installs (via npx skills add) do not get this hook — there, checkpoint proactively on your own per the triggers above.

Recall (history search)

To find past work, search the Checkpoint Log + Archive (not the Current State — that is already in context):

• Full plugin: run /recall <words> (or "${CLAUDE_PLUGIN_ROOT}/scripts/mm-recall" <words>). It returns whole entries containing all the words (case-insensitive; matched per word, not as a phrase), most-recent first.
• Skill-only: grep .claude/memory/CONTEXT.md yourself — scan the ### entries under ## Checkpoint Log and ## Archive for the user's terms and report the matches.