recall

name	recall
description	Use when the user references prior work, asks "what did we decide about X", or starts a topic that may have been explored before — loads `.claude/memory/index.md`, picks the matching tier-2 topic file, and brings the relevant facts into the current turn. Performs a tier-aware retrieval (pointer → topic → optional transcript) instead of dumping everything.
when_to_use	Reach for this at the start of a session that continues prior work, when a user phrase matches a known memory topic, or when grounding a decision against past context. Do NOT use for writing or updating memories — use `/remember` instead. Do NOT use for auditing memory hygiene — use `/memory-index` instead.
disable-model-invocation	true
allowed-tools	["Read","Glob","Grep","Bash"]
logical	matched topic content surfaced ranked by relevance × staleness × confidence, with conflicts resolved toward the higher-trust entry; source path and staleness label included

/recall — Retrieve from Memory

Retrieval Process

Step 1: Read the index

Read .claude/memory/index.md to see all stored pointers. This is Tier 1 — always small.

Step 2: Load relevant topics

For each pointer that matches the current need, read the full topic file from .claude/memory/topics/<slug>.md. This is Tier 2 — loaded on demand.

Step 3: Search transcripts (rare)

If Tier 1 and Tier 2 don't have what you need, search session transcripts with grep:

grep -r "keyword" ~/.claude/projects/*/

This is Tier 3 — never load whole files, only grep for specific terms.

Retrieval Ranking and Staleness Protocol

Every topic file has a Last verified: date and an optional Confidence: field (high | medium | low; omitted ⇒ medium). When multiple topics match a query, order them by:

composite_trust = relevance × staleness_weight × confidence_weight

Where the weights derive from what the topic file already carries:

Age of Last verified:	staleness_weight
< 7 days	1.0
7–30 days	0.7
> 30 days	0.4

Confidence: value	confidence_weight
high	1.0
medium (or omitted)	0.7
low	0.4

This is a ranking convention for Claude to apply by reading the files, not a computed score. (§4.2, Table 1 — arXiv:2605.26112)

Conflict resolution: when two topic files address the same need, prefer the one with the higher composite_trust score and surface the other as superseded-but-recorded (e.g. "Older entry also found: — lower confidence/staleness weight, shown for completeness").

Display labels (human-facing presentation, unchanged):

• < 7 days old: Present as current knowledge
• 7-30 days old: Present as "Previously noted (may be outdated)"
• > 30 days old: Present as "Noted on — verify before acting on this"

Before recommending from memory:

• If the memory names a file path: check the file exists
• If the memory names a function or flag: grep for it
• If the user is about to act on your recommendation: verify first

"The memory says X exists" is not the same as "X exists now."

Output Format

When presenting recalled memories:

Previously noted (<date>):
- <memory content>
- <memory content>

⚠ Verify before acting — memory may be stale.