Memory Palace ยท FastAPI + PostgreSQL + pgvector

Palace Memory Flow

SecondBrain follows the MemPalace idea: keep original knowledge, auto-organize it into navigable rooms, and retrieve with traceable room-aware search.

20sync file types
250 KBmax synced file size
3072embedding dimensions
3Palace run phases
01
1Ingest
LIVE
Capture
Everything
Knowledge enters as notes, API artifacts, or synced files, then gets chunked and embedded for durable recall.
Input Paths
๐Ÿ“‚
Filesystem sync source
Control Tower tracks folders/repos, scans every 15 minutes by default, and only accepts safe extensions and paths.
๐Ÿ“
Provider artifact write API
`/palace/provider/artifacts` stores memory with idempotency keys so repeated writes resolve cleanly.
passes into embedding + indexing queue
โœ“
Verbatim retention Raw content is stored as first-class item data before retrieval decisions.
โ„น
Hybrid-ready indexing Each chunk gets a vector plus PostgreSQL full-text index support.
02
2Organize
Auto-Build
Wings & Rooms
SecondBrain uses metadata, tags, categories, and title tokens to auto-create room structure without manual taxonomy setup.
1
Wing inference Rules combine path parts + tags/categories + title tokens
AUTO
2
Room inference Prefers second path segment, then tags/categories, then title fallback
AUTO
3
Stable key + membership routing Room keys prevent duplicates; auto memberships coexist with pinned curation
TRACKED
4
Snapshot + tunnels Room summaries and cross-room links rebuilt from shared tags (top 4 edges)
GRAPH
โœ“ All pass
Run advances through routing, snapshotting, and tunneling.
โœ• Any fail
Run is marked failed and visible in Control Tower for retry.
03
3Retrieve
LIVE
Room-Scoped
Search + Trace
Queries are routed to likely rooms first, expanded through tunnels when confidence is low, then fall back globally when needed.
Execution Flow
๐Ÿงญ
Route query to candidate rooms
Token-overlap scoring picks top rooms and selected wing.
๐Ÿ•ณ๏ธ
Expand through tunnels
Low confidence (<0.5) widens search using neighboring rooms.
๐Ÿ”Ž
Hybrid Postgres retrieval
0.7 vector similarity + 0.3 `ts_rank`, then best chunk per item.
Guardrails
1-20result limit per call
300-86400sallowed sync interval window
Tracerouting/fallback steps always returned
โšก Indexing bannerwarns when room generation is stale