Memory Search

---
name: memory_search
description: Search past conversations and memory files
always_include: true
cli: true
---
# Memory Search

Search across past conversations and memory files using hybrid BM25 + semantic search.

## Usage

Run `istota-skill memory_search --help` (or `istota-skill memory_search <subcommand> --help`) to see the live argument list.

```bash
# Search for relevant memories
istota-skill memory_search search "query text" [--limit 10] [--source-type TYPE]

# Index a specific conversation
istota-skill memory_search index conversation TASK_ID

# Index a specific file, from your own workspace only
istota-skill memory_search index file {workspace}/notes.md [--source-type TYPE]

# Reindex all conversations and memory files
istota-skill memory_search reindex [--lookback-days 90]

# Show indexing stats
istota-skill memory_search stats
```

`index file` reads from your own workspace or your task's own temp directory (`$ISTOTA_DEFERRED_DIR`) and nowhere else. What it indexes comes back out through `search` after this task is over, so a Talk attachment or a file in the channel directory — material somebody else put there — is refused. Copy it into `{workspace}` first if it belongs in your searchable memory.

## Knowledge graph (read-only)

Structured facts (entity-relationship triples) extracted by the nightly sleep cycle. Current facts are loaded into your prompt automatically as "Known facts."

```bash
# Query current facts (optionally filter by subject or predicate)
istota-skill memory_search facts [--subject ENTITY] [--predicate TYPE]

# Query facts as of a past date
istota-skill memory_search facts --as-of 2026-01-15 [--subject ENTITY]

# Get full timeline for an entity (current + historical facts)
istota-skill memory_search timeline ENTITY
```

**To write facts** — `add-fact`, `invalidate`, `delete-fact` — see the `memory` skill. The classification gate there decides when a piece of information goes to USER.md vs the knowledge graph; consult it before writing either.

## Source Types

- `conversation` — past task prompts and results
- `memory_file` — dated memory files from sleep cycle
- `user_memory` — USER.md persistent memory
- `channel_memory` — CHANNEL.md persistent memory

## Output

```json
{
  "status": "ok",
  "query": "project alpha",
  "count": 3,
  "results": [
    {
      "chunk_id": 42,
      "content": "User: What's the status of project alpha?\n\nBot: Project alpha is...",
      "score": 0.032,
      "source_type": "conversation",
      "source_id": "156",
      "bm25_rank": 1,
      "vec_rank": 3
    }
  ]
}
```

## When to Use

**Always search when:**
- User references past conversations, decisions, or agreements
- User mentions preferences or patterns you should already know
- Topic involves a project, person, or system discussed before
- User says "remember," "we talked about," "last time," or similar

**Also search when:**
- Starting work on a project that may have prior context
- User asks about established patterns, conventions, or preferences
- You want to ensure consistency with previous advice or decisions
- A related topic was likely discussed even if user doesn't reference it

**Don't search when:**
- Query is entirely self-contained with no historical dimension
- You already have sufficient context from the current conversation
- Topic is clearly new with no prior discussion possible

### Search tips

- Use specific terms: "project alpha deployment" not "that project"
- Try multiple queries if first results aren't relevant
- Search for both the topic and related concepts
- Check `--source-type memory_file` for curated memories vs raw conversations

## Notes

- Conversations are automatically indexed after completion
- Memory files are indexed after sleep cycle writes them
- Use `reindex` to backfill if memory search was enabled after conversations already occurred
- Results are scoped to the current user — no cross-user leakage
- Falls back to keyword search if semantic search is unavailable