memory

Search and reason over Atris journal history using the RLM pattern (grep first, reason second).

When to Activate

User asks about:

• Past work, decisions, history
• "Remember when...", "How did we...", "Why did we..."
• Patterns, failures, lessons learned
• "What have we tried before?"
• Anything requiring historical context

Journal Locations

atris/logs/YYYY/YYYY-MM-DD.md

Structure of each journal:

• ## Inbox - Raw ideas (I1, I2, ...)
• ## In Progress 🔄 - Active work
• ## Backlog - Deferred work
• ## Notes - Session summaries, brainstorms
• ## Completed ✅ - Finished work (C1, C2, ...)

Search Strategy (RLM Pattern)

Step 1: Grep first (cheap, fast)

# Find keyword matches
grep -r "keyword" atris/logs/ --include="*.md"

# With context
grep -r -C 3 "keyword" atris/logs/ --include="*.md"

# Multiple terms
grep -r -E "auth|login|token" atris/logs/ --include="*.md"

Step 2: If few matches (< 10), read directly

• Use Read tool on matching files
• Synthesize answer yourself

Step 3: If many matches (10+), use subagent

Task(haiku): "Analyze these journal entries and find patterns related to [query]:
[paste relevant grep results]"

Step 4: For complex synthesis

• Chunk results by time period or topic
• Spawn multiple haiku subagents
• Aggregate findings

Example Flows

Simple: "When did we add feature X?"

1. grep -r "feature X" atris/logs/
2. Read the matching file
3. Answer: "Added on 2025-01-02, see C3 in that day's journal"

Medium: "What auth issues have we had?"

1. grep -r -E "auth|login|token|credential" atris/logs/
2. Found 15 matches across 8 files
3. Read the 3 most recent matches
4. Task(haiku): "Categorize these auth-related entries: [entries]"
5. Synthesize into answer

Complex: "Why do reviews keep failing?"

1. grep -r -E "fail|❌|reject|REVIEW" atris/logs/
2. Found 30+ matches
3. Task(haiku): "What are the failure reasons in: [chunk 1]"
4. Task(haiku): "What are the failure reasons in: [chunk 2]"
5. Aggregate: "78% missing tests, 22% outdated MAP.md"

Key Patterns to Search

| Looking for | Grep pattern | |-------------|--------------| | Completed work | Completed\|✅\|C[0-9]+: | | Failures | fail\|❌\|reject\|block | | Decisions | decided\|decision\|chose\|pivot | | Ideas | Inbox\|I[0-9]+:\|idea\|maybe | | Technical debt | debt\|todo\|hack\|fixme\|refactor |

Cost Efficiency

• Grep: Free, instant
• Read: Counts against context, use sparingly
• Task(haiku): ~$0.001, use for semantic analysis
• Task(sonnet): ~$0.01, use only if haiku insufficient

Always grep first. Only escalate to LLM when you need reasoning, not retrieval.