Memory recall

Cross-Session Memory Recall

Announce: "I'm using enggenie:memory-recall to search previous session context."

Overview

Search past work across sessions using claude-mem's MCP tools. Token-efficient 3-layer retrieval: search the index first, filter, then fetch only what matters.

Requires: claude-mem plugin installed. If not installed:

• When invoked by other skills → skip silently, proceed without memory
• When invoked directly by user → show: "enggenie:memory-recall requires the claude-mem plugin. Install with: claude plugin add claude-mem"

3-Layer Workflow

Never fetch full details without filtering first. 10x token savings.

Layer 1: Search - Get Index with IDs

search(query="authentication", limit=20, project="my-project")

Returns a lightweight table (~50-100 tokens per result):

| ID | Time | T | Title | Read |
|----|------|---|-------|------|
| #11131 | 3:48 PM | feature | Added JWT authentication | ~75 |
| #10942 | 2:15 PM | bugfix | Fixed auth token expiration | ~50 |

Parameters:

• query (string) - Search term
• limit (number) - Max results, default 20, max 100
• project (string) - Project name filter
• type (string, optional) - "observations", "sessions", or "prompts"
• obs_type (string, optional) - Comma-separated: bugfix, feature, decision, discovery, change
• dateStart / dateEnd (string, optional) - YYYY-MM-DD or epoch ms
• offset (number, optional) - Skip N results
• orderBy (string, optional) - "date_desc" (default), "date_asc", "relevance"

Layer 2: Timeline - Get Context Around Interesting Results

timeline(anchor=11131, depth_before=3, depth_after=3, project="my-project")

Or find anchor automatically:

timeline(query="authentication", depth_before=3, depth_after=3, project="my-project")

Returns chronologically ordered items around the anchor point.

Layer 3: Fetch - Get Full Details ONLY for Filtered IDs

Review titles from Layer 1 and context from Layer 2. Pick relevant IDs. Discard the rest.

get_observations(ids=[11131, 10942])

Returns complete observation objects (~500-1000 tokens each): title, subtitle, narrative, facts, concepts, files.

Always use get_observations for 2+ observations - single request vs N requests.

Code Exploration (AST-Based)

When exploring code structure, use token-efficient AST tools instead of full file reads:

1. smart_search - Find symbols across codebase (~2-6k tokens vs ~39-59k for Explore agent)
2. smart_outline - Get file structure (~1-2k tokens vs ~12k+ for full Read)
3. smart_unfold - See specific function implementation (~400-2k tokens)

How Other Skills Use Memory

Every skill that references memory does this:

IF memory-recall MCP tools available:
  Search for relevant context
  Use findings in skill logic
ELSE:
  Skip silently - proceed without memory
  No error message, no mention of missing feature

Skills that use memory:

• enggenie:dev-brainstorm - "Have we designed something similar?"
• enggenie:dev-plan - "What patterns did we use last time?"
• enggenie:pm-refine - "Have we built something similar?"
• enggenie:dev-debug - "Have we seen this bug pattern before?"

Token Savings

Operation	Without memory-recall	With memory-recall
Find past work	Read full conversation logs	Search index: ~100 tokens/result
Explore code	Explore agent: ~39-59k tokens	smart_search: ~2-6k tokens
Read file structure	Full Read: ~12k+ tokens	smart_outline: ~1-2k tokens
See one function	Read full file: ~5-10k tokens	smart_unfold: ~400-2k tokens

Examples

Find recent bug fixes:

search(query="bug", type="observations", obs_type="bugfix", limit=20, project="my-project")

Find what happened last week:

search(type="observations", dateStart="2025-11-11", limit=20, project="my-project")

Understand context around a discovery:

timeline(anchor=11131, depth_before=5, depth_after=5, project="my-project")

Batch fetch details:

get_observations(ids=[11131, 10942, 10855], orderBy="date_desc")

When Search Returns No Results

1. Broaden the query - try synonyms, related terms, or the feature area instead of specific implementation details
2. Check recent sessions - use timeline with a recent anchor to see what was discussed recently
3. Skip gracefully - if no relevant results after 2 search attempts, proceed without memory context. Say: "No relevant past context found. Proceeding fresh."

Do not spend more than 2 search attempts. Memory is a shortcut, not a requirement.

Staleness

Memory observations are snapshots in time. Before acting on a recalled decision:

• Architecture decisions (ADRs, tech choices): Verify the decision is still in effect. Check if the code reflects it.
• Bug patterns ("we fixed this by..."): Check if the same fix applies. The codebase may have changed.
• Process decisions ("we agreed to..."): Ask the user if this is still the team convention.

When in doubt, treat memory as a starting point for investigation, not as ground truth.

Recommended Model

Primary: haiku Why: Memory retrieval is about searching indexes and fetching observations. Haiku is fast and efficient for this lookup-heavy work.

This is a recommendation. Ask the user: "Confirm model selection or override?" Do not proceed until the user responds.

---

Entry Condition

None - available anytime. Invoked by other skills as a utility, or directly by user.

Exit Action

Context provided → resume whatever workflow triggered the recall.