How it works
Curator pipeline, resolver, salience scoring. Learn more →
Memory
Memory
Section titled “Memory”Ptah remembers. Long-running work tends to outlive a single context window — preferences shift, decisions accumulate, and “what did we settle on for auth?” becomes a real question. Ptah’s memory subsystem captures durable knowledge from your sessions and surfaces it back to the agent on demand.
The design is Letta-style tiered:
| Tier | Cap (default) | What lives here |
|---|---|---|
| core | 50 | Identity, durable preferences, project invariants |
| recall | 500 | Recent events, decisions, things that happened this week |
| archival | 5000 | Deep history — searchable but rarely surfaced |
A curator runs automatically on every context compaction (driven by the SDK’s PreCompact hook). It reads the conversation, extracts memory drafts, then a resolver merges those drafts against existing memories — inserting, updating, promoting, demoting, or forgetting. Salience scoring decides which memories rise to core and which decay out of archival.
Search is hybrid: BM25 (SQLite FTS5) and vector similarity (sqlite-vec, 384-dim bge-small-en-v1.5 embeddings) are fused with Reciprocal Rank Fusion. If sqlite-vec isn’t available on your platform, search falls back to BM25-only — no setup pain, just slightly less semantic recall.
Memory kinds: fact, preference, event, entity.
What’s in this section
Section titled “What’s in this section”Searching
Hybrid BM25 + vector retrieval, top-K, weighting. Learn more →
Pinning & forgetting
Pin durable facts; soft-delete what no longer applies. Learn more →
Settings
Every tunable, with defaults. Learn more →