Skip to content
smara

Notes from building a memory.

What we shipped, what we measured, and what we got wrong — written the way we'd want another engineering team to write it.

  • Same harness, or it didn't happen

    We ran Mem0 through our own LoCoMo harness — same questions, same models, same judge. 74.7% vs 63.1%, every config fix disclosed, plus a confession of our own.

  • What Waku got right

    A competitor landed on our exact thesis. Instead of a subtweet: the list of what they did better, and what we built about it in a week.

  • Talking to your own memory

    The Stream answers in the first person and cites the exact memories it drew on — including, some mornings, its own dreams.

  • A graph worth walking

    Five upgrades that took our knowledge graph from extraction byproduct to retrieval surface — including one that didn't move the benchmark.

  • The ledger you can replay

    Our marketing site promised a visible episodic ledger before the product had one. Closing that gap became the Provenance tab.

  • A team of agents that remembers

    Shared memory pools with per-key grants, webhooks on memory changes, and content policies — the multi-agent wedge, explained.

  • Auditable dreams: why our nightly consolidation leaves receipts

    Every memory system is shipping 'dreaming' this year. Ours is the only one you can audit edit-by-edit the morning after.

  • Four LoCoMo runs, three lessons

    68.4 → 74.2 → 74.1 → 74.7 — and one number we chose not to publish. What benchmarking a live memory service honestly actually looks like.

  • Your memory should travel

    One memory across ChatGPT, Claude, and Gemini — arrive with years of history, leave with everything. Anti-lock-in as an architecture, not a promise.

Influences

Ideas and products that made Smara better, named. Good competitors keep a category honest; good ideas deserve citations.

  • LoCoMo (Maharana et al., ACL 2024) — the benchmark discipline our whole measurement culture grew around.
  • Ebbinghaus's forgetting curve — the century-old shape behind our decay model: memory strength as time-since-rehearsal.
  • GraphRAG (Microsoft Research) — community detection over entity graphs, the seed of our community summaries.
  • Mem0 — made agent memory a category; our same-harness comparison exists because their numbers were worth comparing against.
  • Waku Memory — set the bar for capture consent UX and one-command onboarding — smara-capture and npx smara setup answer it.
  • Stripe's webhook signatures — the t=…,v1=… signing scheme our memory-change webhooks borrow outright.