Memory Model¶
Hermit's memory model is built around a simple idea:
memory should not be an ungoverned place where assertions quietly become authority
That is why Hermit separates working state, beliefs, and durable memory records.
The Three Layers¶
Working State¶
Working state is bounded execution state for active work.
It should be:
- task-scoped
- size-bounded
- shaped for continuation and planning
- easy to supersede
Working state is not the same thing as durable memory.
Beliefs¶
Beliefs are revisable working truths.
A belief represents what the system currently treats as true enough to reason with inside or near a task boundary. Beliefs can cite evidence, carry confidence, and later be invalidated or superseded.
In practical terms, beliefs help Hermit avoid two bad choices:
- pretending every claim is durable memory
- pretending every claim should disappear with the current turn
Memory Records¶
Memory records are durable cross-task knowledge.
They are for facts and conventions that should survive beyond the current task, but only under governance rules such as:
- evidence references
- scope
- retention class
- expiration
- invalidation or supersession
Why Evidence Matters¶
Memory without provenance becomes hidden system prompt.
Hermit's direction is that durable memory promotion should cite evidence and remain inspectable later. This makes memory less like a bag of sticky notes and more like a governed knowledge layer.
Scope And Retention¶
The current codebase already contains memory governance logic for:
- scope matching
- retention classes
- expiration
- static injection eligibility
- retrieval eligibility
- supersession logic
Examples of scope kinds include:
- global
- conversation
- workspace
- entity
This matters because not every memory should follow the system everywhere.
Promotion And Retrieval¶
The practical lifecycle is:
- working activity produces claims
- some claims become beliefs
- some beliefs become durable memory records
- retrieval logic decides what can re-enter context
- invalidation and supersession keep memory from silently rotting
The key design choice is that promotion is not just "the model thought this sounded useful."
What Exists Today¶
Safe claims:
- the repository already defines
BeliefRecordandMemoryRecord - the repository already has memory governance services
- current logic already covers classification, scope, retention, expiry, and supersession paths
Careful claims:
- the memory model is materially real, but still evolving
- evidence-bound memory is a shipped direction with partial implementation, not a fully settled final system
Why This Matters¶
Many agent systems say they have memory. Fewer say what kind of memory, under what scope, with what evidence, and with what invalidation rule.
Hermit's memory model matters because durable work needs:
- memory that can be trusted enough to use
- memory that can still be challenged
- memory that does not silently outrank evidence