Is this marketing speak or real research?

The features follow concrete, documented findings from neuro- and cognitive science (consolidation, reward-prediction-error replay, Complementary Learning Systems, prospective memory, reflection). We transfer mechanisms, not metaphors - and explicitly claim no consciousness.

Why do not Mem0/Zep/Letta have this?

Most optimize extraction and retrieval but accumulate at the core. Active forgetting, prospective memory and an auditable self-model are open gaps we built on purpose.

Does the forgetting curve lose my data?

No. The value only affects the order of search results - it deletes and filters nothing. An entry you have not needed for two years is still there and still findable, it just no longer sits in front of what you use daily. Deletion happens only when you ask for it.

How is "surprise" measured?

By the semantic distance to the most similar entry that already existed at write time. If nothing comparable is found, the information was new. If an entry explicitly contradicts earlier knowledge, a premium is added. The value feeds into ranking and into the order of the nightly distillation.

Recallbase / Blog / Brain science

What we built into AI memory from brain science

Most AI memory systems only accumulate. The brain does the opposite: it forgets on purpose, consolidates selectively and remembers at the right moment. We rebuilt a few of these principles - here is which and why.

A glowing ring of connected nodes - consolidated memory

Forgetting is not a failure, it is a function

In the brain, forgetting is an active, energy-costly process - the prefrontal cortex actively inhibits goal-irrelevant memories, and that is exactly what enables generalization. AI memory systems, by contrast, mostly accumulate, and their retrieval precision degrades with volume. Recallbase actively mutes outdated or refuted knowledge (instead of just flagging it) and keeps it out of context - still reachable for history, but no longer poisoning decisions.

The forgetting curve, as a formula

Since July 2026 every memory also carries a stability in days. How present it is follows a forgetting curve from spaced-repetition research: R = exp(-age / stability). Every real retrieval raises that stability (14 days become 23, then 38 ...), capped at roughly 13 months. Knowledge you use fades more slowly, knowledge you never touch drifts down the ranking - so the index stays signal-dense after months instead of silting up.

Deliberately damped: the value shifts the order, it never filters. Rarely used specialist knowledge should still be findable half a year later - it simply is not at the top anymore. Actively hiding something stays reserved for suppression, which needs a documented reason.

Consolidation is selective, not blanket

At night the hippocampus does not replay everything - it prefers the surprising, replay follows the reward-prediction-error. Recallbase weights origin and trust (your statement beats a web find) and lets salient, correcting knowledge survive preferentially. And it dreams at night: an LLM distills crashed sessions and recurring patterns.

Surprise as a number

For us that is not a metaphor but a value on the record. On write, Recallbase measures the distance to the nearest existing entry: if nothing comparable is in memory, the information was new and scores high. Anything that explicitly refutes existing knowledge gets an extra premium - a violated expectation is the strongest learning moment, in the brain and here.

That score lifts the entry in ranking and pulls it preferentially into the nightly distillation (prioritized replay). And much like tag-and-capture in the hippocampus, salience bleeds onto the direct neighbours in the knowledge graph, damped by half: whatever happened around a surprise gets consolidated with it. Unlike decay, this value does not fade - the moment of surprise was what it was.

From episodic to schema (Complementary Learning Systems)

The hippocampus learns fast and episodic, the neocortex gradually takes over schema-like knowledge. Recallbase mirrors this: sessions are episodic memory, the knowledge graph is semantic. What is confirmed across several sessions gets promoted to cross-project rules - many single cases become one rule for the next project.

Prospective memory: remembering before being asked

Humans hold intentions with triggers ("when I pass the store, buy milk"). This is precisely where, per benchmarks, all large language models and memory products fail. Recallbase stores intentions with a trigger condition and surfaces them actively at the right moment - push instead of pull.

A self-model you can audit

Autobiographical memory shapes identity. Recallbase derives a sober working profile from actual behavior - working style, strengths, recurring mistakes, every claim tied to a source session. When the aspiration contradicts the behavior ("claims TDD but often skips tests"), it is flagged as drift. Not an invented persona block, but traceable.

Staying honest

None of this is "consciousness" or a "soul" - it is pragmatic transfer of well-documented memory mechanisms onto a database. But these very principles are missing from most systems that treat memory as "load more context". Try it free.

A memory that thinks like one

Consolidate selectively, forget on purpose, remember at the right moment. 1 project free.

Start for free