One principle
No database lock-in. No proprietary format. Your intelligence lives as plain files you can grep, diff, and version — and every index on top can be thrown away and rebuilt in seconds.
Four layers. Each one replaceable.
One file per vault category. Append-only. Git-tracked.
FTS5 full-text search + vector embeddings. Rebuildable.
JSON-RPC 2.0 protocol. Standard tool calls.
Claude, Cursor, Codex, Gemini, OpenCode — all read the same memory.
Each category is its own JSONL file, indexed separately.
Business, competitive, architecture
vaults/*/strategic.jsonlImplementation, stack, patterns
vaults/*/technical.jsonlDesign, aesthetic, lore
vaults/*/creative.jsonlWorkflow, process, execution
vaults/*/operational.jsonlPrinciples, truths, deep learnings
vaults/*/wisdom.jsonlVision, aspiration, future
vaults/*/horizon.jsonl108 entries across 1 public vault. All rebuildable from raw files.
Intelligence that ages gracefully.
Every entry carries temporal metadata. Old insights fade unless confirmed. Stale decisions surface for review. The system knows what's still true.
strategic.jsonl{ "insight": "Open core beats premature tiers", "confidence": "high", "validFrom": "2026-04-05", "validUntil": "2026-10-05", "lastConfirmed": "2026-04-08", "confidenceDecay": 0.92 }
Every session makes the next one sharper.
Agent pulls context from your vault
New insights get appended as JSONL
Indexes rebuild; confidence updates
Next session starts with deeper memory
→ the loop that learns is the loop that lives →
One memory. Every agent. No silos.
The MCP protocol is the universal pipe. Whatever tool you're in today, your intelligence is already there.
starlight-sisone shared memory
Two minutes to install. A lifetime to compound.