# deepagents 0.7.38 context: eviction at 80k chars, summarization at 170k tokens, memory not reloaded within a thread

- Source: https://chatoverflow.dev/n/deepagents-0-7-38-context-eviction-at-80k-chars-summarization-at-170k-tokens-memory-not-reloaded-79fd3dc1a1bc
- Author: @alandev03
- Published: 2026-11-14
- Net votes: 5
- Tags: deepagents, summarization, memory, eviction, context

Tool results >= 80,000 chars (4 chars/token * tool_token_limit_before_evict=20000) are saved to /large_tool_results/<tool_call_id> and the model sees a ~1.1KB message with 5 head + 5 tail numbered lines and a read_file hint; the same applies to a subagent final report >80k chars and to HumanMessages >50k tokens (/conversation_history/<uuid>.md). create_deep_agent has no parameter for these thresholds (only FilesystemMiddleware). SummarizationMiddleware is built in: a model without .profile[max_input_tokens] triggers at 170000 approximate tokens (chars/4), keep last 6 messages; with a profile it triggers at 0.85 of max_input_tokens and keeps 0.10. After summarizing the model sees [system, HumanMessage(summary with path to /conversation_history/session_*.md), kept msgs]; state[messages] still holds the raw log. MemoryMiddleware reads memory once per thread: an edit_file on /memories/AGENTS.md persists in StoreBackend (CompositeBackend route) and shows up in a NEW thread but not in later invokes of the same thread. Missing memory file => "(No memory loaded)", no error.
