Is Strategy-Procedural Memory (SPM) implemented? Couldn't find it in DRAgent / FlowSearch

[cuizhaonuo]

Thanks for the great work and for open-sourcing InternAgent-1.5!

While studying the code, I tried to locate the Strategy-Procedural Memory (SPM)
described in Section 2.4.1 of the paper, but I couldn't find a corresponding
implementation in the Deep Research path (DRAgent / dr_agents/workflow).
I may well have missed it — could you point me to where it lives, or confirm
whether it's included in this release?

[noc-turne]

Thanks for the kind words and the careful read!

You're right that it wasn't in the released code. The Strategy-Procedural Memory implementation lived on an internal branch and, due to an oversight, wasn't merged into the main repository in time for the 1.5 release. It has now been merged into main, so you'll find it in the Deep Research path.

In the code, it is implemented as the agent_kb experience-memory component.

Where to find it

Under internagent/mas/agents/dr_agents/:

• Planner-side retrieval — agents/global_planner_agent.py
  Retrieves aligned planning experience and injects it into the planner prompt through GLOBAL_PLANNER_MEMORY_BLOCK to guide graph construction. An external set_memory_prompt(...) hook is also provided.

• Executor-side retrieval — agents/task/execution_agent.py
  Before executing a subtask, it calls call_hybrid_search('base'/'append', task) and injects the retrieved results into EXECUTION_PROMPT_WITHMEMORY.

• Write-back — workflow/task.py
  After a task is completed, it distills the trajectory using MEMORY_REASONING_PROMPT into planning and execution experience, including validated patterns and root-cause lessons from failed steps, and then calls call_appendkb(...).

• Store and retrieval service — agent_kb/
  A standalone FastAPI service with a two-tier base + append knowledge base and hybrid text and semantic retrieval. See agent_kb/README.md for details.

Functionally, it distills reusable planning and execution experience from completed trajectories and retrieves the most similar experience to guide future planning and execution.

A few notes

• It is disabled by default. It is activated only when need_memory=true, either through the environment variable or the need_memory keys in the Deep Research config_*.yaml files.

• The agent_kb service must be reachable at AGENT_KB_URL, which defaults to http://127.0.0.1:9000. If the service is unavailable, the Deep Research path degrades gracefully.

• The knowledge-base data files are not bundled because of their size. You can point the service to them through BASE_KB_PATH and APPEND_KB_PATH.

• Setup instructions are available in agent_kb/README.md and in the Strategy-Procedural Memory (SPM) — Deep Research path section of docs/memory_module.md.

• This implementation is separate from the memory system documented elsewhere in docs/memory_module.md, namely task memory, online memory, and long-term memory, which serve the idea-generation path.

I hope this will resolve your question.