recall

Files (1)

• SKILL.md

SKILL.md (first lines)

# Recalling Past Decisions and Context

**BEFORE answering ANY question about past decisions, architecture, constraints, or project history, you MUST check the Brain's memory.**

## Quick Start

Not sure which tool to use? Call `memory_decision` first:

```
memory_decision(query="<user's question>", session_id="<session>")
```

It returns a `recommended_tool` field telling you exactly which tool to call next. If confidence is low, skip the recall.

## Mandatory Recall Triggers

| When You See This | You MUST Call This |
|---|---|
| Questions about past decisions | `current_truth(query="<decision topic>")` |
| "Why did we choose X?" | `current_truth(query="decision about X")` |
| Architecture or design questions | `current_truth(query="<architecture topic>")` |
| Multi-entity or broad context needs | `compile_context(query="<broad topic>")` |
| "Tell me everything about X" | `compile_context(query="X")` |
| Not sure which tool to use | `memory_decision(query="<question>")` |
| Process or procedure questions | `recall_procedures(query="<process>")` |
| "What went wrong last time?" | `recall_relevant_lessons(query="<topic>")` |
| "What happened this week?" | `reconstruct_timeline(limit=20)` |
| "What changed since yesterday?" | `diff_since(since_iso="<ISO timestamp>")` |

**Do NOT answer from your context window alone.** The Brain's graph is more reliable than your training data for project-specific facts.

## Tool Reference

| Tool | When to Use |
|---|---|
| `current_truth` | Single-topic recall: "why did we choose X?", "what's the constraint on Y?" |
| `compile_context` | Multi-entity queries: "tell me everything about the auth system", broad context needs |
| `memory_decision` | Not sure which tool? Call this first — it recommends the right retrieval tool |
| `recall_procedures` | "How do we deploy?", "What's the process for X?" — procedural knowledge |
| `recall_relevant_lessons` | "What went wrong last time?", "Any lessons about X?" — past outcomes |
| `reconstruct_timeline` | "What happened this week?", "When did we decide X?" — temporal queries |
| `analogical_search` | "Have we done something like this before?" — cross-project patterns |
| `explore_graph` | Browse entity connections: "What's related to X?" |
| `diff_since` | "What changed since yesterday?" — recent changes |

## Scoping

**ALWAYS use scope 'both' when the question involves:**
- Multi-project decisions or patterns
- Constraints that span multiple services or domains
- When you're unsure which scope applies

## Anti-Bloat Rules

- Use top 3 results unless exhaustive review is specifically needed.
- Summarize compactly — memory informs your answer, it IS NOT the answer.
- Do not paste raw memory output. Summarize in 2-3 sentences.
- If results include a `bloat_warning`, mention to the user that the conversation is getting long.

## How to Use Results

- Results are ranked by relevance and confidence
- High pathway_strength = frequently accessed, well-established knowledge
- Items marked `confidence_low` are tentative — flag the uncertainty to the user
- The Brain's graph is more reliable than your context window for historical facts
- If recall returns nothing relevant, say so — don't fabricate context

From engramist/hippocampy, Apache-2.0.

More from engramist/hippocampy