TelloDB

Standard vector stores treat every retrieved chunk as equally valid regardless of age, which breaks the moment user state evolves — addresses change, preferences flip, ownership changes. TelloDB is a memory engine that addresses this by pairing hybrid search (HNSW vector index plus BM25 full-text) with a temporal truth layer that tags older facts as stale when newer ones contradict them. The ingestion pipeline distills raw chat input into structured facts rather than storing raw conversational text, keeping the memory lattice compact and queryable. The workflow is: ingest a natural-language fact, let the distillation kernel extract structured meaning, and query against a clean profile rather than a bloated transcript store.

The differentiating feature is fact supersession combined with deterministic aggregation. When a user states they moved or bought something new, the engine resolves the conflict — marking the prior state stale and committing the new fact as active — rather than leaving contradiction resolution to the LLM at query time. The Metric Vault layer handles numeric and count-based queries (“how many cars do I own?”) through a built-in execution layer before the prompt reaches the LLM, which the vendor positions as a fix for known benchmark failures in pure vector-based memory.

TelloDB fits teams building personalization layers, stateful assistants, or multi-model pipelines where session state must survive backend switches. It ships as a single Rust binary with a self-host path and a one-click hosted deployment option, making it a candidate for air-gapped or data-sovereign deployments via its documented proxy gateway. Where it breaks: at v0.1.0, the tool is explicitly a memory substrate — it does not provide agent task planning, tool-calling scaffolding, or workflow branching. Teams that need those capabilities alongside persistent memory will be composing TelloDB with a separate agent framework and maintaining that integration themselves.