Sieve — Agent Memory Curation Dashboard

📋 Live LLM System Prompt Context (Namespace: default)

Ready for Context Window

- [DECISION] Actually, after review we switched the primary database from PostgreSQL to MongoDB 7.0 for document schemas. - [TASK] TODO: Provision staging cluster environment before Friday's investor demo.

2 Active In-Prompt Items · ~140 Tokens ✓ Pruned 1 Stale Fact (42% Prompt Token Bloat Eliminated)

How Sieve Works in Real Environments

Everything happening in the visual dashboard above maps directly to 3 lines of code in your Python agent.

Zero Vector Drift · SQLite WAL

from sieve import SieveMemory

# 1. Initialize local memory layer (SQLite WAL mode + FTS5 BM25)
memory = SieveMemory("production.db")

# 2. In your agent turn loop: remember what the user says
memory.remember("Switched primary database from PostgreSQL to MongoDB", namespace="tenant_42")

# 3. Inject clean, non-conflicting, single-source-of-truth facts into LLM system prompt
clean_context = memory.get_context(namespace="tenant_42")
system_prompt = f"You are a helpful assistant.\nVerified Context:\n{clean_context}"

STAGE 01 / INGESTION

Microsecond Scoring

Memories are categorized in ~0.19ms without GPU, scoring importance and actionable readiness before touching LLM tokens.

STAGE 02 / PAIRWISE SWEEPS

Contradiction Resolution

FTS5 BM25 pairs related memories. Newer assertions supersede older ones with direct pointer attribution (superseded_by).

STAGE 03 / SAFEGUARDED PRUNING

Clean Prompt Context

Questions never overwrite decisions (100% precision). Pinned directives stay permanent invariants. Stale items are evicted.

🔑 API Key Authentication

How Sieve Works in Real Environments