Reflect is slow issue

[wong218]

my world facts more than 200
Observations more than 90
then Reflect exceeded the 10-second limit, resulting in failure.
Is this a bug? Or—how should I configure it to make it more responsive and efficient, while also saving tokens?
i using minimax 2.7 do it before

Root Cause Analysis

The reason Reflect is slow and token-heavy is that Hindsight's reflect agent is internally a multi-round iterative LLM agent:

• iteration 1: search_observations → returns all 70 entries in full (5.4s)
• iteration 2: recall → vector retrieval for more facts (7.6s)
• iteration 3: recall → another retrieval round (7.9s)
• LLM call 1 (agent_1): analyzes observations (5.2s)
• LLM call 2 (agent_2): analyzes recall results (12.2s)
• LLM call 3 (agent_3): synthesizes (10.9s)
• LLM call 4 (agent_4): generates final answer (30.7s)

─────────────────────────────────────
Total: 80 seconds, 64K input tokens

[nicoloboschi]

what's your environment? 6 seconds for the recall itself is a lot

Minimax 2.7 might be slow itself but the other timings are concerning too