Revisioned Memory

Tentarc memory is revisioned structured state, not a mutable blob.

Data model idea

Each write appends a new revision record:

{
  "workspaceId": "ws_prod",
  "sessionId": "s_42",
  "agentId": "session",
  "revision": 18,
  "status": "ok|fallback",
  "summary": {},
  "meta": {}
}

Read path selects latest valid revision per (sessionId, agentId).

Why revision beats overwrite

• preserves lineage (what changed, when, why)
• supports replay/rollback/fork workflows
• makes partial failure explicit instead of destructive

Fallback semantics

When generation/query fails, runtime can persist a fallback revision:

• revision still increments
• record carries failure metadata
• system remains monotonic and debuggable

No silent memory drop.

Compaction principle

Because revision logs are append-heavy, compaction should:

• preserve latest effective revision
• preserve critical audit metadata
• avoid collapsing unresolved failure history

Compaction is storage optimization, not semantic rewrite.