AuthorsNishant Gaurav, Adit Akarsh, Tejas Ravishankar, Manoj Bajaj
TL;DR
CodeMem uses a Dynamic MCP layer plus a persistent procedural memory bank to turn sandboxed Python workflows into reproducible skills, reaching 96% correctness with Gemini 3 Full.
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
CodeMem targets probabilistic instability where agents follow different trajectories for the same task, wasting planning and debugging time every run.
This instability breaks repetitive workflows like Outlook to OneDrive orchestration, causing trajectory divergence, context collapse, and unreliable execution for multi-step tool chains.
HOW IT WORKS
CodeMem centers on search_functions, load_functions, write_todos, execute_code, and register_skill to discover tools, plan tasks, execute code, and freeze successful logic.
Think of CodeMem as an LLM architect designing scripts, a sandbox CPU executing them, and a procedural memory bank acting like a versioned code library instead of a fragile prompt.
This KEY_MECHANISM lets CodeMem move complex loops and filters into deterministic Python, enabling infinite-depth workflows and reproducible skills that plain context windows and prompt edits cannot guarantee.
DIAGRAM
This diagram shows how CodeMem handles the Outlook OneDrive Bridge case study from discovery to sandbox execution and skill registration.
DIAGRAM
This diagram shows how CodeMem is evaluated on 25 multi step tasks using Gemini 2.5 Flash as an LLM as a Judge.
PROCESS
Discovery and Planning
CodeMem uses search_functions and load_functions to discover Dynamic MCP tools, then write_todos to externalize a Pending to Completed checklist for the task.
Script Generation
CodeMem authors Python using execute_code in a sandbox, composing loops and conditionals that orchestrate tools like Outlook and OneDrive deterministically.
Execution and Self Correction
CodeMem runs the script, inspects stdout and stderr, and updates write_todos to recover state, iterating until the workflow matches user constraints.
Skill Registration and Reuse
CodeMem calls register_skill to freeze the validated agent_main function into procedural memory, enabling future runs to skip planning and coding.
KEY CONTRIBUTIONS
CodeMem Architecture
CodeMem combines Dynamic MCP search_functions and load_functions with execute_code and register_skill to turn ephemeral scripts into deterministic procedural memory for agents.
Procedural Memory via Frozen Code
CodeMem treats procedural memory as versioned Python functions, avoiding brittle prompt updates and ensuring deterministic execution of workflows like the Outlook OneDrive Bridge.
Experimental Benchmark on 25 Tasks
CodeMem evaluates Gemini 3 Full, Claude 4.5 Sonnet, and GPT-5 Chat on 25 multi step tasks, reporting 96% minimum correctness and 100.48s P50 latency for Gemini 3 Full.
RESULTS
Correctness Min
96%
+28 percentage points over GPT-5 Chat
Avg Calls
7.00
Gemini 3 Full vs 2.80 for GPT-5 Chat
P50 Latency
100.48s
Longer latency than 14.75s GPT-5 Chat due to deeper iteration
Total Tokens
2.02M
Token budget for Gemini 3 Full trajectories on 25 tasks
Table 2 reports CodeMem performance on a 25 task benchmark of multi step agentic workflows. The 96% minimum correctness for Gemini 3 Full shows that CodeMem can reliably complete complex tool chains when allowed to iterate about seven times per task.
BENCHMARK
Table 2 reports CodeMem performance on a 25 task benchmark of multi step agentic workflows. The 96% minimum correctness for Gemini 3 Full shows that CodeMem can reliably complete complex tool chains when allowed to iterate about seven times per task.
BENCHMARK
Correctness Min across models when using CodeMem on 25 multi step tasks.
KEY INSIGHT
GPT-5 Chat shows the lowest P50 latency at 14.75s and only 0.49M tokens, yet its minimum correctness is just 68% on CodeMem tasks.
This is surprising because faster, cheaper runs usually seem better, but CodeMem reveals that shallow trajectories with 2.80 calls miss edge cases that higher reasoning models reliably handle.
WHY IT MATTERS
CodeMem unlocks reproducible, deterministic workflows where agents architect Python pipelines once, then reuse them as procedural memory without re planning or re coding.
Builders can now treat LLMs as workflow designers that crystallize skills into code, enabling Zapier like reliability for complex multi tool automations that were previously brittle and stochastic.
Nikhil Verma
· 2026
Focus Agent adds start_focus, complete_focus, a persistent Knowledge block, and an optimized Persistent Bash plus String-Replace Editor scaffold to actively compress context during long software-engineering tasks. On five hard SWE-bench Lite instances against a Baseline ReAct agent, Focus Agent achieves 22.7% token reduction (14.9M → 11.5M) while matching 3/5 = 60% task success.
Xingyu Lyu, Jianfeng He et al.
· 2026
ADAM combines Anchor extraction, Distribution estimation, Anchor selection, and Query generation to adaptively probe agent memory via an auxiliary generator and entropy based selection. On the EHRAgent benchmark with Llama2-7b-chat, ADAM reaches EQ=77 and ASR=1.00, compared to MEXTRA’s EQ=44 and ASR=0.89.
Shannan Yan, Jingchen Ni et al.
· 2026
AdaMem organizes dialogue history into Working Memory, Episodic Memory, Persona Memory, and Graph Memory coordinated by a Memory Agent, Research Agent, and Working Agent. On LoCoMo with GPT-4.1-mini, AdaMem achieves 44.65 F1 overall, beating the best baseline LangMem at 41.76 F1 by +2.89.
Answers use this explainer on Memory Papers.
Checking…