Skill: bastion/life-log

Objective

Maintain a semantic history of interactions per persona, allowing Bastion to:

1. Remember past conversations without user needing to repeat context
2. Search most relevant records to enrich responses (RAG)
3. Analyze usage patterns per persona over time
4. Detect inactive personas for proactive suggestions

CLI Commands

IMPORTANT: CLI Command When running inside Bastion, invoke these operations through the declared tool interface. Do not simulate tool results.

from skills.life_log.factory import Settings, create_adapter
from datetime import datetime, timezone

# Instantiate via factory (reads DB_STRATEGY from environment)
settings = Settings.from_env()
adapter = create_adapter(settings)

# Log an interaction
interaction_id = await adapter.log_interaction(
    persona="tech-lead",
    intent="code-review",
    tools=["github-integration", "code-review-helper"],
    embedding=[0.1, 0.2, ...],   # generated by the configured LLM
    timestamp=datetime.now(tz=timezone.utc),
)

# Semantic search (RAG)
similar = await adapter.search_similar(
    query_embedding=[0.1, 0.2, ...],
    persona="tech-lead",          # None to search across all personas
    limit=5,
    threshold=0.65,
)

# Summary of last 7 days
summary = await adapter.get_persona_summary(
    persona="tech-lead",
    days=7,
)

Required Fields for Each Record

Semantic Search

Threshold and Limit

• default threshold: 0.65 — only results with cosine similarity ≥ 0.65
• default limit: 5 — maximum 5 results per search
• Results ordered by descending similarity

Embedding Generation

Embeddings are generated by the provider configured in Bastion. The skill doesn't generate embeddings directly — receives the vector already calculated by orchestrator.

Context Injection in Prompt

The orchestrator uses search_similar before assembling each response:

1. Generate embedding from user input
2. Call search_similar(query_embedding, persona=active_persona, limit=5, threshold=0.65)
3. Inject returned records as historical context into the system prompt
4. Respond with enriched context

DB_STRATEGY Pattern

The skill uses the Protocol/Adapter (hexagonal) pattern to decouple business logic from the persistence backend.

Switching from SQLite to Supabase

Just change one variable in .env:

# Default — local SQLite, zero external dependency
DB_STRATEGY=sqlite
SQLITE_PATH=db/life-log.db

# Optional — Supabase (PostgreSQL + pgvector)
DB_STRATEGY=supabase
SUPABASE_URL=https://your-project.supabase.co
SUPABASE_KEY=your-anon-or-service-key

No skill code needs to change. The factory (create_adapter) instantiates the correct adapter based on DB_STRATEGY.

Available Adapters

SQLite Persistence

Database Location

db/life-log.db   ← created automatically on first write

The db/ directory is created automatically if it doesn't exist.

Schema

CREATE TABLE interactions (
    id        TEXT PRIMARY KEY,   -- UUID v4
    persona   TEXT NOT NULL,      -- persona slug
    intent    TEXT NOT NULL,      -- executed intent
    tools     TEXT NOT NULL,      -- JSON array of tool names
    embedding BLOB NOT NULL,      -- float32 little-endian
    timestamp TEXT NOT NULL       -- ISO 8601 UTC
);

CREATE INDEX idx_interactions_persona   ON interactions (persona);
CREATE INDEX idx_interactions_timestamp ON interactions (timestamp);

sqlite-vec

When the sqlite-vec extension is installed, it is loaded automatically for native vector search. If not available, the adapter falls back to pure-Python cosine similarity — works offline with no native dependencies.

Architecture (Hexagonal)

LifeLogProtocol (port)
    ├── log_interaction(persona, intent, tools, embedding, timestamp) → str
    ├── search_similar(query_embedding, persona, limit, threshold) → list[InteractionRecord]
    └── get_persona_summary(persona, days) → list[InteractionRecord]

SQLiteLifeLogAdapter (default adapter)
    ├── Creates db/life-log.db automatically
    ├── Uses sqlite-vec when available
    └── Falls back to pure-Python cosine similarity

SupabaseLifeLogAdapter (optional adapter)
    ├── Uses supabase-py + pgvector
    └── Stub — requires full implementation for production

Settings + create_adapter() (factory)
    └── Instantiates the correct adapter based on DB_STRATEGY

When This Skill is Triggered

1. Interaction Logging (after each response)

The orchestrator calls log_interaction at the end of each relevant interaction:

await adapter.log_interaction(
    persona=active_persona.slug,
    intent=detected_intent,
    tools=tools_used,
    embedding=await llm.embed(user_input),
    timestamp=datetime.now(tz=timezone.utc),
)

2. Semantic Search (before each response)

The orchestrator enriches context before responding:

context = await adapter.search_similar(
    query_embedding=await llm.embed(user_input),
    persona=active_persona.slug,
    limit=5,
    threshold=0.65,
)
# Inject context into system prompt

3. Weekly Review (HEARTBEAT — every Monday at 9am)

summary = await adapter.get_persona_summary(persona=slug, days=7)
# Analyze patterns, suggest weight adjustments

4. Inactivity Detection (HEARTBEAT — every 6h)

for persona in all_personas:
    recent = await adapter.get_persona_summary(persona=persona.slug, days=3)
    if not recent:
        # Generate resumption suggestion

Edge Cases

Database doesn't exist on first read

The SQLiteLifeLogAdapter creates the file and schema automatically before any operation. No manual initialization required.

Embedding with different dimension

search_similar compares embeddings using cosine similarity. If vectors have different dimensions, a ValueError is raised. Always use the same embedding model.

Threshold too high (no results)

If no record reaches the threshold, search_similar returns an empty list. The orchestrator must handle this case and respond without historical context.

Persona with no history

get_persona_summary returns an empty list for personas with no interactions in the period. This is not an error — it indicates a new or inactive persona.

Output Example

{
  "id": "a1b2c3d4-e5f6-7890-abcd-ef1234567890",
  "persona": "tech-lead",
  "intent": "code-review",
  "tools": ["github-integration", "code-review-helper"],
  "timestamp": "2024-01-15T10:30:00Z"
}