Why Flat Vector Retrieval Fails on Long-Term Agent Memory (And How Bi-Temporal SVO Modeling Fixes It)

When building autonomous AI agents, developers often start with a flat vector database (like ChromaDB or Pinecone) to store conversation history. However, as agent sessions scale, flat vector search faces a fundamental limitation: it cannot encode time.

The Core Problem: Flat Vectors Don't Know "When"

If a user says "I live in New York" on Monday, and then "I moved to San Francisco" on Friday, a standard vector search for "Where does the user live?" will retrieve both statements with high cosine similarity. The agent receives conflicting facts and hallucinates.

The Solution: Bi-Temporal Dual-Calendar Memory (Smriti)

To solve this, Smriti uses a bi-temporal architecture in PostgreSQL:

1. SVO Event Extraction: Turns are parsed into Subject-Verb-Object (SVO) assertions with explicit valid_from and valid_to timestamps.
2. Zero-LLM Fact Supersession: When a new fact arrives, an in-process statistical engine invalidates the prior valid_to timestamp without needing slow, expensive LLM reflection passes.
3. Point-in-Time Queries: Queries evaluate valid_to IS NULL for active truths, keeping the context window 100% clean.

Check out the live API and benchmarks at smriti-kaal.vercel.app.