GGraph: An Optional Graph & Semantic Acceleration Layer (Preview)

An optional shadow accelerator that sits on top of Postgres to make associative recall faster on large graphs — and never takes over correctness.

While your memory is small, everything is fast. A few hundred or a few thousand nodes, and a local database (or your own Postgres) pulls connected memories back in milliseconds. But Corsoul's associative memory keeps growing with lived experience: the web of associations between memories gets denser, following those associations has more paths to walk, and semantic recall has more candidates to compare. Past a certain size, the act of finding related memories can start to become the bottleneck.

GGraph exists for exactly that moment — an optional acceleration layer. It doesn't ask you to migrate anywhere; it just makes the graph you already have run faster.

What it is: not a new database, a shadow

Let's be clear about the one thing that matters most first: GGraph is not a new database, and you don't migrate any data into it.

Corsoul's durable source of truth is always your local database or your own Postgres. GGraph sits on top of that as a "shadow" — it maintains an accelerated copy of the associative and semantic structures purely to serve the two heaviest kinds of reads. How you write, how you store, where your data lives — none of that changes. GGraph only opens a fast lane on the read side.

The design is deliberately conservative. The one thing a memory system can't get wrong is the facts themselves, so Corsoul's truth stays on the proven database foundation. GGraph only touches how quickly you can find what's already there — never what is true.

What it accelerates: associative and semantic recall

Corsoul's associative memory relies on two motions to gather what's relevant back together.

The first is recall that spreads along associations: starting from a cue and lighting up connected memories layer by layer, following the links between them, with the most relevant surfacing most strongly. On a small corpus this is instant; when the associations grow many and dense, the number to traverse climbs quickly.

The second is semantic recall: comparing the current context against a large body of stored memories to surface the ones closest in meaning. The more candidates there are, the more this step costs.

GGraph targets precisely these two — keeping associative and semantic recall responsive at scale, so that associative memory still keeps pace with the conversation even after your corpus has grown. What it accelerates is speed; what it returns is the same set of memories.

The safety net: correctness is always guaranteed by fallback

This is the most important — and most trust-worthy — property of GGraph: its correctness does not depend on itself.

Any time GGraph errors, becomes unavailable, or returns something unexpected, that path transparently falls back to the underlying database, and the proven foundation computes the result directly. Your agent never sees a failure — only a recall that may be slightly slower but is correct. In other words, the accelerator is a bonus, not a load-bearing wall. Remove it entirely and Corsoul's memory behavior stays complete and correct; it's just less fast on large corpora.

Because the safety net is shaped this way, turning GGraph on is a low-risk decision: the worst case is falling back to the performance you already had.

When to consider it (and its boundaries)

Honestly, most users don't need GGraph at first. While your corpus is small to moderate, a local database is plenty fast on its own, and you likely won't feel the difference an accelerator makes.

The signal that it's actually worth considering is concrete: your corpus keeps growing and associative queries start to become the bottleneck — spreading slows down, recall begins to drag on how immediate a response feels. That's when GGraph becomes an optional, non-destructive upgrade.

And its status deserves the same honesty: GGraph is currently Experimental / Preview, not GA. We don't claim it has been validated at large-scale production. Its value today rests exactly on the safety net above — you can try it against your own scale and workload, watch how it behaves, and never bet the correctness of your memory on it. That's why we're comfortable putting it on the table now without overselling it.

Closing

Corsoul's underlying philosophy is consistent throughout: local-first, free to start, light up deeper capability only when you need it. Objective memory — remember, recall, forget, and prospective intents — is fully usable offline on the free tier. When lived experience grows into a large graph that needs association and pattern, the personality engine takes over learning and consolidation; and when that graph gets big enough that even reading it should be faster, GGraph is the optional, fallback-protected, honestly-labeled-Preview way to accelerate.

Start free, let memory grow naturally, and accelerate when you need to.

Memory becomes experience. Experience becomes a self.