Long-term memory: an operation that learns should remember why.
Most systems keep a log of what happened. Almost none keep a record of what was decided, why, and how it turned out. AnantState does, with who made each decision, so future cases can be read against real precedent.
In plain termsA memory of decisions and their results, not just of events.
Four kinds of memory
| Memory | Contents |
|---|---|
| Entity history | The state trajectory of every entity, with the events that acted on it. |
| Decisions | Every recommendation, applied action, override and decline, with its reason. |
| Outcomes | What actually happened after each decision: the part almost nobody records. |
| Learnings | Human-attributed notes and conclusions attached to an entity, case or pattern. |
The third is the one that matters most and is most often missing. A record of decisions without outcomes teaches nothing; the pairing is what turns an operation’s history into an asset.
It changes the answers, not just the archive
Memory is governed, not accumulated
Because a memory you cannot trust is worse than none.
| Property | How |
|---|---|
| Attributed | Every decision, override and note carries its actor. |
| Tenant-scoped | Isolated by the same data-layer control as everything else. No cross-tenant leakage. |
| Retained by policy | Retention windows are configured, not unlimited. |
| Queryable | A first-class surface, not a log file to be searched. |
| Explainable | A conclusion drawn from memory states which prior cases it used. |
Retention matters. Unbounded memory accumulates stale regimes: a pattern from three years ago can outvote the current one. Policies define how far back memory remains influential, and that policy is a business decision with an audit trail.
Memory and learning are not the same thing
| Long-term memory | Learning | |
|---|---|---|
| What it is | A record of what happened, was decided and resulted | A change in how the platform predicts |
| Where it lives | Governed, inspectable, attributable | In the model’s own parameters |
| How it changes | Continuously, as events and decisions arrive | At training time, through a gate |
| Who controls it | Retention and access policy | Promotion policy |
| Can it be audited? | Yes: it is data | Yes: via the model record and its gates |
Memory changes what the platform knows; learning changes what it predicts. Conflating them is how a system ends up quietly rewriting its own behavior with no record of why.
Start remembering the outcome, not just the action
We will show how a decision receipt is recorded in one step and how precedent reads on the reference world.