Skip to content

The agent that grows with you

fix(agent): the summary prompt and the system prompt leave the token memo too

kshitijk4poor26e2a21d024b
diff --git a/agent/context_compressor_telemetry.py b/agent/context_compressor_telemetry.py
index 233e9e1cfc7..d8b9c4d92fe 100644
--- a/agent/context_compressor_telemetry.py
+++ b/agent/context_compressor_telemetry.py
@@ -10,6 +10,7 @@ import uuid
from typing import Any
from agent.model_metadata import estimate_messages_tokens_rough
+from agent.model_metadata_token_memo import forget_message_token_memo
def _safe_int(value: Any) -> int | None:
@@ -78,6 +79,8 @@ class CompressionTelemetryMixin:
if not isinstance(telemetry, dict):
return
telemetry["aux_prompt_tokens"] = estimate_messages_tokens_rough(prompt_messages)
+ # The summary prompt is built fresh per attempt and never estimated again; don't let the memo pin it.
+ forget_message_token_memo(prompt_messages)
telemetry["aux_output_reservation"] = _safe_int(max_tokens)
if aux_provider:
telemetry["aux_provider"] = aux_provider
diff --git a/run_agent.py b/run_agent.py
index 9ed59accb2f..a2220dc5bad 100644
--- a/run_agent.py
+++ b/run_agent.py
@@ -954,9 +954,10 @@ class AIAgent(
self._db_flush_scan_prefix = None
self._streamed_assistant_text_parts = []
if isinstance(session_messages, list):
- # The token-estimate memo pins message strings; a closed conversation's can never hit again.
+ # The token-estimate memo pins message strings; a closed conversation's can never hit again. The
+ # system prompt reaches the memo only through api_messages, never through the history.
from agent.model_metadata_token_memo import forget_message_token_memo
- _quietly(forget_message_token_memo, session_messages)
+ _quietly(forget_message_token_memo, (session_messages, getattr(self, "_cached_system_prompt", None)))
_quietly(self._trim_process_memory)
_quietly(self._finalize_owned_session_row)
diff --git a/tests/agent/test_token_memo_release_on_close.py b/tests/agent/test_token_memo_release_on_close.py
index 5381dc7d817..b80b40b1acf 100644
--- a/tests/agent/test_token_memo_release_on_close.py
+++ b/tests/agent/test_token_memo_release_on_close.py
@@ -74,3 +74,29 @@ def test_compression_evicts_the_bodies_the_summary_replaced(tmp_path, monkeypatc
assert id(dropped) not in _pinned_ids()
assert id(kept) in _pinned_ids() # still live in the compressed history
+
+
+def test_close_evicts_the_system_prompt_the_history_never_holds():
+ system_prompt = "a long per-session system prompt body " * 2000
+ estimate_messages_tokens_rough([{"role": "system", "content": system_prompt}])
+ agent = _closing_agent([{"role": "user", "content": "hi"}])
+ agent._cached_system_prompt = system_prompt
+ assert id(system_prompt) in _pinned_ids()
+
+ agent.close()
+
+ assert id(system_prompt) not in _pinned_ids()
+
+
+def test_compression_telemetry_does_not_pin_the_one_shot_summary_prompt():
+ from agent.context_compressor_telemetry import CompressionTelemetryMixin
+
+ recorder = CompressionTelemetryMixin()
+ recorder._active_compression_telemetry = {"effective_aux_context": None}
+ prompt = "summarize this serialized transcript " * 20000
+
+ recorder._record_aux_compression_call(prompt_messages=[{"role": "user", "content": prompt}],
+ max_tokens=None, duration_ms=1)
+
+ assert recorder._active_compression_telemetry["aux_prompt_tokens"] > 0
+ assert id(prompt) not in _pinned_ids()