Smara

Smara gives Claude durable, structured memory that persists across conversations — so an agent that learned something about you, your project, or your customers yesterday still knows it tomorrow. Unlike a notes file, Smara is real memory infrastructure. What agents write is consolidated in the background into durable facts and a knowledge graph of entities and relationships, with full provenance. Retrieval is hybrid — vector search, graph traversal, and recency-weighted ranking packed into a token budget with a single tool call. Memory strengthens when it is used and fades when it is not: importance-weighted, usage-reinforced decay keeps what matters and forgets session noise, the way human memory does. Seven tools cover the full lifecycle: get_context assembles everything relevant to a turn (scratchpad, recent turns, facts, graph relationships, past episodes) in one call; capture_event and append_turn record what happens; remember_fact and search_facts read and write durable knowledge; search_events answers "what happened?"; patch_scratchpad keeps short-term working state. Everything is governed: memory is scoped to a tenant you choose at connect time, every change is recorded in a bi-temporal audit trail (nothing is silently overwritten), and you can browse and delete any fact or entity from the developer portal at smara.dev. Deliberate deletion is immediate; nothing about your data is used for training. Built for people building agents: the same memory is also a plain REST API with an OpenAPI spec, so what your Claude connector writes, your own applications can read — one memory across every surface. Measured, not promised: 68.4% on the LoCoMo long-term-memory benchmark, run against this live service with a published, reproducible methodology.