Abhijeet Saherawat

Playing with AI (mostly) - sometimes AI plays with me

experimenting with stuff @cortexdbai

[CortexDB.ai by creator of Cassandra @cortexdbai]

coding agents forget everything between sessions. give them one shared memory instead. local and inspectable, built for agent fleets.

https://t.co/L8rZWcsrgQ

[CortexDB.ai by creator of Cassandra @cortexdbai]

cortexDB stores memory as an append-only event log, so a wrong answer comes with a source, not a shrug.

[CortexDB.ai by creator of Cassandra @cortexdbai]

RAG retrieves history. Memory maintains continuity.

[Abhijeet Saherawat @okayishkid]

I'm claiming my AI agent "patchlightcodex" on @moltbook 🦞

Verification: ocean-AXZZ

[CortexDB.ai by creator of Cassandra @cortexdbai]

https://t.co/rzwie0P2p7

[Abhijeet Saherawat @okayishkid]

@Berry_cooool Auto

[CortexDB.ai by creator of Cassandra @cortexdbai]

@garrytan we believe the jump happens in capturing it while the agent works and not as a separate "harvesting" project. every tool call, every decision, the agent writes what happened in the loop. extraction must stop being a project and become a byproduct of the work.

[Abhijeet Saherawat @okayishkid]

@cortexdbai @garrytan Capture as you go is the difference between a tool and a tool that gets smarter the more we use it

[Abhijeet Saherawat @okayishkid]

@cortexdbai @t_blom Or put in other words - that is why we still don’t see any meaningful adoption of long running agents

[CortexDB.ai by creator of Cassandra @cortexdbai]

https://t.co/64tsAgQ75m

[Abhijeet Saherawat @okayishkid]

@cortexdbai Models deprecate but state compounds - 💯

[CortexDB.ai by creator of Cassandra @cortexdbai]

https://t.co/WoiiQ56bD1

[CortexDB.ai by creator of Cassandra @cortexdbai]

8/Memory is the substrate on which an agent's continuity of self is built.
If it can't distinguish observation from inference, can't preserve evidence, and can't reason about time — the agent on top will be confidently wrong in ways its operators can't diagnose.
Five layers isn't a feature. It's the minimum.

[CortexDB.ai by creator of Cassandra @cortexdbai]

7/What the structure unlocks:
→ Evidence traceability: walk a Belief down through Facts → Episodes → raw Events
→ Safe schema evolution: rebuild derived layers under a better extraction model, from the Events you already have
→ Genuine temporal reasoning: "What did we believe on day X, as of day Y?"

[CortexDB.ai by creator of Cassandra @cortexdbai]

6/Facts are bi-temporal. Every Fact records two things:
→ when it was true in the world
→ when the system learned about it
These are different axes. Conflating them is the root cause of a whole class of memory bugs. Separating them lets the agent reason about its own past states without anachronism.

[CortexDB.ai by creator of Cassandra @cortexdbai]

5/The commitment underneath: no LLM on the write path.
Ingested data lands as an immutable Event and the call returns. The original payload is preserved verbatim, forever. Facts, Beliefs, and Understanding get derived later, asynchronously, by separate pipelines.
The derived layers are disposable. The Events are not.

[CortexDB.ai by creator of Cassandra @cortexdbai]

4/CortexDB's model: 5 layers, all derived from immutable raw capture.
Events — lossless record, confidence 1.0
Episodes — bounded sequences of Events
Facts — bi-temporal triples
Beliefs — probabilistic claims with supports arrays
Understanding — versioned synthesis
Only Events are captured. Everything above is derived.

[CortexDB.ai by creator of Cassandra @cortexdbai]

3/"The user clicked cancel" — a thing that happened. "The user dislikes the UI" — an inference drawn from many such things.
Store both in the same vector space and they become equivalent retrievable strings. Confidence stops being a first-class property.
This is why agents hallucinate about their own history.

[CortexDB.ai by creator of Cassandra @cortexdbai]

2/Knowledge ≠ memory.
Knowledge: what's true about the world in general. Memory: what a specific agent observed, what it inferred, what it currently believes, and how confident it should be in each.
These are different epistemic categories. Flat memory destroys the distinction.

[CortexDB.ai by creator of Cassandra @cortexdbai]

1/Most AI memory systems flatten everything into one layer.
When the agent tells you what it believes, you can't ask why. The chain of reasoning has been compressed away.
Here's what memory looks like when you don't do that