"Agent memory" usually means a vector database: embed everything the user said, query by similarity, paste the top matches into the prompt. It's a useful trick, but it isn't memory. It's a lookup table that never learns, never forgets correctly, and can't tell you what was true last month versus today. An agent built on it doesn't get smarter the longer you run it — it just accumulates more haystack to search.
The name Eident ic is deliberate: an agent without memory has no identity. We think real memory needs four things working together.
Plain vector recall has no concept of time. If a user was on the starter plan in March and the team plan in June, both sentences sit in the index with equal weight, and the model picks whichever embeds closer. That's how agents confidently tell you yesterday's truth.
Eidentic stores facts in a temporal knowledge graph where each fact carries a validity interval. New information supersedes the old without deleting it: the agent can answer "what plan are they on now" and "what plan were they on in April" from the same store, and contradictions resolve instead of piling up. Memory that can't reason about time isn't memory — it's a cache.
People don't store a transcript of every conversation; they keep a running summary and revise it. Eidentic gives agents self-editing memory blocks — compact, structured notes the agent rewrites as it learns — plus passive extraction that pulls salient facts out of every turn automatically. You don't write ingestion pipelines or decide what to remember; the agent maintains its own working memory.
If memory only ever grows, retrieval gets slower and noisier over time — the opposite of improving. Eidentic runs sleep-time consolidation: between sessions it compresses and merges what was learned, so the next session starts knowing more without a larger prompt. This is the step that makes memory self-improving rather than merely cumulative. Lexical and vector retrieval each miss things the other catches, so Eidentic fuses both with reciprocal-rank fusion and returns results with citations. An answer drawn from memory can point at the session and the fact it came from — which matters the moment an agent does anything consequential.
Wiring a model to a tool loop is a weekend. Memory that stays correct as it grows — across contradictions, across time, without ballooning the prompt — is the part teams underestimate and then rebuild three times. It's also what separates a demo from an agent you'd put in front of real users for months.
An agent without memory has no identity. Eidentic gives agents theirs — and keeps it honest as the world changes.
This isn't theoretical: on long histories it's measurably better and cheaper than stuffing everything into context (the benchmarks). If you're building something an agent has to remember, start with the docs or the source.