Proxy behavior & safety
SPM makes independent decisions for recall, capture, and history shortening. Seeing recall activity does not imply that compression is allowed.
| Request shape | Recall | History shortening | Continuity state |
|---|---|---|---|
| Ordinary long chat/history with searchable older exchanges | Yes | May commit removal or fold | committed or fold receipt |
| Short/new history | Yes | Nothing to remove | not_needed |
| Memory unavailable or source coverage unproven | May be empty/degraded | No | bypassed_* |
Responses with previous_response_id |
Tail only | No | bypassed_protected_context |
| Reasoning/tool state that cannot be safely rewritten | Yes, when allowed | No | bypassed_protected_context |
shadow mode |
Yes, when allowed | Measured only | shadow_only |
Testing the two paths
For eligible compression, use a normal multi-turn chat request whose older exchanges have already become searchable memory. Verify lower forwarded tokens, a committed continuity state, and (when a fold is selected) a fold economics receipt.
For provider-managed Responses state, send a request with previous_response_id.
SPM preserves provider history and may add only recalled memory at the tail. A
large request with a small positive memory addition is intentional, not evidence
that SPM was skipped.
Token counters are SPM estimates used for routing, compression, and receipts;
provider usage may differ. recalled_tokens > 0 means memory was injected; it
does not mean older history was removed.
Provider reasoning, thinking, tool-call arguments, tool-result pairing, and provider-managed history are never rewritten for savings. If coverage, readiness, or evidence is insufficient, SPM forwards the complete request.