A memory engine that behaves like infrastructure: deterministic, temporal, and verifiable. A system of record your agents, and your auditors, can rely on.
Most memory layers paraphrase what was said, overwrite what changed, and can’t show their work. That’s fine for a demo — and a liability in production.
Extraction rewrites messages into lossy summaries. When a claim is challenged, there’s nothing exact to point back to.
New facts clobber old ones in place, so you can never reconstruct what the agent believed at a past moment.
A soft-delete flag is a pinky-swear. Nothing an auditor can verify without trusting your database.
When agents touch health, money, or people’s records, “probably deleted” isn’t an answer. Shomei was built from the ground up so that every claim it makes about memory can be checked, layer by layer, by someone who doesn’t trust us.
A patient asks what your assistant knows about them, or asks it to forget. Both requests resolve to evidence, not a support ticket.
When a regulator asks why the agent said what it said, you replay the exact memory state it saw, not an approximation of it.
An HR case closes, a customer leaves, a court order lands. Inside the memory we govern, erasure leaves no residue: not the words, not their derivatives, not a hash that could confirm what was there. The receipt proves it.
Facts change. Shomei’s graph is bi-temporal: when a new fact contradicts an old one, the old fact retires with a date instead of being overwritten. Ask about the past and the graph still answers.
Most memory products store an LLM’s paraphrase. Shomei stores the exact span of the original message. If the bytes are not in the source, nothing is stored: hallucinated memories are ruled out by construction.
The leaders are genuinely good at recall. The gap is everything after: reproduce it, trace it, and prove that what you deleted left no residue behind.
| PRODUCT | FOCUS | DETERMINISTIC? | SOURCE-LINKED? | PROVABLE DELETE? |
|---|---|---|---|---|
| mem0 | recall, add-only memory | no · LLM extraction | no | flag / append |
| Zep / Graphiti | temporal knowledge graph | no · LLM graph build | no | shared edges linger |
| Letta (MemGPT) | agent-managed memory | no · agent rewrites | no gate | overwrite |
| Shomei | deterministic temporal graph + proof | byte-identical replay | verbatim spans | erased + receipt |
Competitor defaults change, verify from each vendor’s current docs before quoting externally.
One engine takes your agents’ conversations in and hands receipts out. Everything between is deterministic and inspectable.
Extract facts as verbatim spans of the source message, with offsets, dates, and a subject.
Place each fact on the bi-temporal graph; contradictions retire old facts with a date.
Every retrieval re-checks policy at use time, so a restriction or deletion holds everywhere, instantly.
Recall, restrict, hold, export, or erase: each emits a signed receipt our open verifier checks offline.
Drop it into an agent in an afternoon: in-process SDK, HTTP service, or MCP server. Runs in your cloud, on your keys.
Deterministic, temporal, provable memory for AI agents. Show me the proof.